Re: An question for FICLONERANGE ioctl

[Sidong Yang <realwakka@gmail.com>]

On Mon, Sep 06, 2021 at 02:30:21PM +0800, Qu Wenruo wrote:
> 
> 
> On 2021/9/6 下午1:57, Sidong Yang wrote:
> > On Mon, Sep 06, 2021 at 09:13:06AM +0800, Qu Wenruo wrote:
> > > 
> > > 
> > > On 2021/9/5 下午8:14, Sidong Yang wrote:
> > > > Hi, All.
> > > > I've tried to handle btrfs-progs issue.
> > > > (https://github.com/kdave/btrfs-progs/issues/396)
> > > > 
> > > > And I tested some code like below.
> > > > 
> > > > src_fd = open(src_path, O_RDONLY);
> > > > if (src_fd < 0) {
> > > > 	error("cannot open src path %s", src_path);
> > > > 	return 1;
> > > > }
> > > > 
> > > > dest_fd = open(dest_path, O_WRONLY|O_CREAT, 0666);
> > > > if (dest_fd < 0) {
> > > >       close(src_fd);
> > > >       error("cannot open dest path %s", dest_path);
> > > >       return 1;
> > > > }
> > > > 
> > > > range.src_fd = src_fd;
> > > > range.src_offset = src_offset;
> > > > range.src_length = length;
> > > > range.dest_offset = dest_offset;
> > > 
> > > Mind to give an example of the value?
> > It was src_offset = 0, src_length = 10, dest_offset = 0.
> Oh, that's the case.
> 
> > 
> > > 
> > > One quick hint to the invalid arguments is:
> > > 
> > > - Range alignment
> > >    The src/dst offset must be aligned to the block size of the
> > >    filesystem.
> > >    For btrfs, the sectorsize is currently the same as page size,
> > >    thus both src/dest and length must be aligned to 4K for x86.
> > 
> > I think it's because of this. I set too small value. It works with
> > length 4K. If reflink cmd in btrfs-progs exists, Users should set the
> > length aligned?
> 
> Reflink/clone/dedupe all work based on block size.
> 
> Currently all these major files systems (ext*/xfs/btrfs/f2fs) in linux
> are block filesystems, which means block size is their minimal unit of
> read/write.
> 
> For data stored on-disk, they are all aligned to block size and smaller
> ranges are padded to meet block alignment.
> 
> So is such ioctls (and things like direct IO).
> 
> Thus no matter what the user-space wrapper is, the kernel ioctl still
> requires block size alignment.

I understood that all datas are aligned to block size.
Thanks a lot for detailed description!

Thanks,
Sidong
> 
> Thanks,
> Qu
> 
> > 
> > Thanks,
> > Sidong
> > > 
> > > Thus a more detailed example can be much better for us to understand the
> > > problem.
> > > 
> > > Thanks,
> > > Qu
> > > > 
> > > > ret = ioctl(dest_fd, FICLONERANGE, &range);
> > > > 
> > > > And this ioctl call failed with error code invalid arguments when length!=0.
> > > > I tried to understand FICLONERANGE man page but I think there is no clue
> > > > about this. I traced kernel code and found out it goes fail in
> > > > generic_remap_checks(). There is an condition checks if req_count is
> > > > correct size and it makes test code fails.
> > > > 
> > > > I don't know about this condition but it seems that it can be passed for
> > > > setting REMAP_FILE_CAN_SHORTEN. Is there any way to setting remap_flags
> > > > in FICLONERANGE ioctl call?
> > > > 
> > > > Also it would be pleased that if you provide some documentation about
> > > > this.
> > > > 
> > > > Sorry for writing without thinking deeply.
> > > > 
> > > > Thanks,
> > > > Sidong
> > > > 
> >