[LTP] [PATCH] syscalls/copy_file_range02.c: Compatible with new and old kernels

[Yang Xu]

on 2019/06/27 16:39, Amir Goldstein wrote:
> On Thu, Jun 27, 2019 at 11:03 AM Yang Xu<xuyang2018.jy@cn.fujitsu.com>  wrote:
>> on 2019/05/31 21:02, Amir Goldstein wrote:
>>
>>> On Fri, May 31, 2019 at 3:03 PM Li Wang<liwang@redhat.com>   wrote:
>>>>
>>>> On Fri, May 31, 2019 at 6:15 PM Xiao Yang<yangx.jy@cn.fujitsu.com>   wrote:
>>>>> On 2019/05/31 16:44, Jinhui huang wrote:
>>>>>> On new kernels, copy_file_range() returned EISDIR when copyed contents
>>>>>> to directory, but on old kernels, it returned EBADF, we should accept
>>>>>> EBADF to be compatible with new and old kernels.
>>>>>>
>>>>>> The patch as follows:
>>>>>> commit 11cbfb10775a ("vfs: deny copy_file_range() for non regular files")
>>>>> Hi,
>>>>>
>>>>>   From description of commit, I wonder if we can add more tests for some
>>>>> non regular files(e.g. block, pipe)?
>>>> I have no objection on this. But, is there really make sense to test some more non regular files which not being mentioned by Linux Manual Page?
>>>>
>>> FYI, more changes to copy_file_range checks are in the works:
>>> https://lore.kernel.org/linux-fsdevel/20190526061100.21761-1-amir73il@gmail.com/T/#me34d4363449118bd3b2ec8421d282a77e9a7d2d1
>> Hi Amir
>>
>>       Meet again.  We can increase ltp copy_file_range02 coverage include( swapfile ->ETXTBUSY,  immutable file->EPERM) as same as xfstests generic/553[4].
>> Also the two other checks(overlaping and offset wrap).  I am glad to do it.
> That would be great!

Hi Amir

I have tried it.  swapfile and immutable file has no problem, but when I try to reproduce EINVAL(same file overlaping) without xfs_io, I got EFAULT error.
It look likes we must depend on the new xfs_io feature patch.  Right?

ps: If it must xfs_io command, I think we can not check this situation because we should only check by copy_file_range syscall.
what do you think about it?

another question:
I have seen copy_file_range manpage,  it says fd_out data can be overwriting, but I got EFAULT when I use the following code.

open(src_path, O_RDWR|O_CREAT, 0644) = fd_copy
open(copy_path, O_RDWR|O_CREAT, 0644) = fd_src
SAFE_WRITE(1, fd_src,  CONTENT,  CONTSIZE);
SAFE_WRITE(1, fd_copy,  CONTENT,  CONTSIZE);
copy_file_range(fd_src,0, fd_copy, CONTSIZE/4, CONTSIZE/2, 0) = -1 EFAULT

fs/read_write.c copy_file_range function copy_from_user reports this error. If off_out or off_in isn't equal to 0, the error occurs.
  
---------------------
ret= -EFAULT;
....
if (off_out) {
                copy_from_user(&pos_out, off_out, sizeof(loff_t));

                         goto out;
              }
----------------------

Is it a bug or I miss something?

Thanks
Yang Xu

>
>> PS: Why we don't have test for overlaping and offset wrap check on xfstests? Or, I miss it?
> The bounds check test was posted to xfstests:
> https://marc.info/?l=linux-xfs&m=155947929219690&w=2
>
> But because the test requires a new feature from xfs_io:
> https://marc.info/?l=linux-xfs&m=156152984109774&w=2
>
> I recommended that xfstests maintainer will hold off merging the test,
> before the
> required change is at least in xfsprogs for-next branch.
>
> Thanks,
> Amir.
>
>
>