2013-03-07 04:05:12

by Qian Cai

Subject: change of behavior for madvise in 3.9-rc1

Bisecting indicated that this commit,
1998cc048901109a29924380b8e91bc049b32951
mm: make madvise(MADV_WILLNEED) support swap file prefetch

Caused an LTP test failure,
http://goo.gl/1FVPy

madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
madvise02 5 TFAIL : madvise succeeded unexpectedly

While it passed without the above commit
madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
madvise02 5 TPASS : failed as expected: TEST_ERRNO=EBADF(9): Bad file descriptor

CAI Qian

2013-03-07 04:39:48

by Shaohua Li

Subject: Re: change of behavior for madvise in 3.9-rc1

On Wed, Mar 06, 2013 at 11:05:04PM -0500, CAI Qian wrote:
> Bisecting indicated that this commit,
> 1998cc048901109a29924380b8e91bc049b32951
> mm: make madvise(MADV_WILLNEED) support swap file prefetch
>
> Caused an LTP test failure,
> http://goo.gl/1FVPy
>
> madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> madvise02 5 TFAIL : madvise succeeded unexpectedly
>
> While it passed without the above commit
> madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> madvise02 5 TPASS : failed as expected: TEST_ERRNO=EBADF(9): Bad file descriptor

I thought this is expected behavior. madvise(MADV_WILLNEED) to anonymous memory
doesn't return -EBADF now, as now we support swap prefretch.

Thanks,
Shaohua

2013-03-07 18:50:31

by Hugh Dickins

Subject: Re: change of behavior for madvise in 3.9-rc1

On Thu, 7 Mar 2013, Shaohua Li wrote:
> On Wed, Mar 06, 2013 at 11:05:04PM -0500, CAI Qian wrote:
> > Bisecting indicated that this commit,
> > 1998cc048901109a29924380b8e91bc049b32951
> > mm: make madvise(MADV_WILLNEED) support swap file prefetch
> >
> > Caused an LTP test failure,
> > http://goo.gl/1FVPy
> >
> > madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> > madvise02 5 TFAIL : madvise succeeded unexpectedly
> >
> > While it passed without the above commit
> > madvise02 1 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 2 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 3 TPASS : failed as expected: TEST_ERRNO=EINVAL(22): Invalid argument
> > madvise02 4 TPASS : failed as expected: TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> > madvise02 5 TPASS : failed as expected: TEST_ERRNO=EBADF(9): Bad file descriptor
>
> I thought this is expected behavior. madvise(MADV_WILLNEED) to anonymous memory
> doesn't return -EBADF now, as now we support swap prefretch.

I agree with Shaohua: although the kernel strives for back-compatibility
with userspace, I don't think that goes so far as to tell an arbitrary LTP
test that it has failed, once the kernel has been enhanced to support new
functionality. We could never add or extend system calls if that were so.

Hugh

2013-03-08 02:05:28

by Qian Cai

Subject: Re: change of behavior for madvise in 3.9-rc1

----- Original Message -----
> From: "Hugh Dickins" <[email protected]>
> To: "Shaohua Li" <[email protected]>
> Cc: "CAI Qian" <[email protected]>, "linux-mm" <[email protected]>, "linux-kernel" <[email protected]>,
> "Rik van Riel" <[email protected]>, "Sasha Levin" <[email protected]>, "Andrew Morton"
> <[email protected]>, "Linus Torvalds" <[email protected]>
> Sent: Friday, March 8, 2013 2:49:48 AM
> Subject: Re: change of behavior for madvise in 3.9-rc1
>
> On Thu, 7 Mar 2013, Shaohua Li wrote:
> > On Wed, Mar 06, 2013 at 11:05:04PM -0500, CAI Qian wrote:
> > > Bisecting indicated that this commit,
> > > 1998cc048901109a29924380b8e91bc049b32951
> > > mm: make madvise(MADV_WILLNEED) support swap file prefetch
> > >
> > > Caused an LTP test failure,
> > > http://goo.gl/1FVPy
> > >
> > > madvise02 1 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 2 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 3 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 4 TPASS : failed as expected:
> > > TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> > > madvise02 5 TFAIL : madvise succeeded unexpectedly
> > >
> > > While it passed without the above commit
> > > madvise02 1 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 2 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 3 TPASS : failed as expected:
> > > TEST_ERRNO=EINVAL(22): Invalid argument
> > > madvise02 4 TPASS : failed as expected:
> > > TEST_ERRNO=ENOMEM(12): Cannot allocate memory
> > > madvise02 5 TPASS : failed as expected:
> > > TEST_ERRNO=EBADF(9): Bad file descriptor
> >
> > I thought this is expected behavior. madvise(MADV_WILLNEED) to
> > anonymous memory
> > doesn't return -EBADF now, as now we support swap prefretch.
>
> I agree with Shaohua: although the kernel strives for
> back-compatibility
> with userspace, I don't think that goes so far as to tell an
> arbitrary LTP
> test that it has failed, once the kernel has been enhanced to support
> new
> functionality. We could never add or extend system calls if that
> were so.
Thanks for looking this. We will try to fix the LTP test instead.
>
> Hugh
>