Re: copy_file_range and user space tools to do copy fastest

From: Steve French
Date: Sat Apr 28 2018 - 01:27:08 EST


On Sat, Apr 28, 2018 at 12:18 AM, Andreas Dilger <adilger@xxxxxxxxx> wrote:
> On Apr 27, 2018, at 5:41 PM, Eric Biggers <ebiggers3@xxxxxxxxx> wrote:
>>
>> On Fri, Apr 27, 2018 at 01:45:40PM -0600, Andreas Dilger wrote:
>>> On Apr 27, 2018, at 12:25 PM, Steve French <smfrench@xxxxxxxxx> wrote:
>>>>
>>>> Are there any user space tools (other than our test tools and xfs_io
>>>> etc.) that support copy_file_range? Looks like at least cp and rsync
>>>> and dd don't. That syscall which now has been around a couple years,
>>>> and was reminded about at the LSF/MM summit a few days ago, presumably
>>>> is the 'best' way to copy a file fast since it tries all the
>>>> mechanisms (reflink etc.) in order.
>>>>
>>>> Since copy_file_range syscall can be 100x or more faster for network
>>>> file systems than the alternative, was surprised when I noticed that
>>>> cp and rsync didn't support it. It doesn't look like rsync even
>>>> supports reflink either(although presumably if you call
>>>> copy_file_range you don't have to worry about that), and reads/writes
>>>> are 8K. See copy_file() in rsync/util.c
>>>>
>>>> In the cp command it looks like it can call the FICLONE IOCTL (see
>>>> clone_file() in coreutils/src/copy.c) but doesn't call the expected
>>>> "copy_file_range" syscall.
>>>>
>>>> In the dd command it doesn't call either - see dd_copy in corutils/src/dd.c
>>>>
>>>> Since it can be 100x or more faster in some cases to call
>>>> copy_file_range than do reads/writes back and forth to do a copy
>>>> (especially if network or clustered backend or cloud), what tools are
>>>> the best to recommend?
>>>>
>>>> Would rsync or cp be likely to take patches to call the standard
>>>> "copy_file_range" syscall
>>>> (http://man7.org/linux/man-pages/man2/copy_file_range.2.html)?
>>>> Presumably not if it has been two+ years ... but would be interested
>>>> what copy tools to recommend to use instead.
>>>
>>> I would start with submitting a patch to coreutils, if you can figure
>>> out that code enough to do so (I find it quite opaque). Since it has
>>> been in the kernel for a while already, it should be acceptable to the
>>> upstream coreutils maintainers to use this interface. Doubly so if you
>>> include some benchmarks with CIFS/NFS clients avoiding network overhead
>>> during the copy.
>>>
>>
>> For cp (coreutils), apparently there was a concern that copy_file_range()
>> expands holes; see the thread at
>> https://lists.gnu.org/archive/html/bug-coreutils/2016-09/msg00020.html.
>> Though, I'd think it could just be used on non-holes only. And I don't think
>> the size_t type of 'len' is a problem either, since it's the copy length, not
>> the file size. You just call it multiple times if the file is larger.
>
> I think cp is already using SEEK_HOLE/SEEK_DATA and/or FIEMAP to determine
> the mapped and sparse segments of the file, so it should be practical to
> use copy_file_range() in conjunction with these to copy only the allocated
> parts of the file.

For the case where clone/reflink or copy_file_range is supported - is
there any reason to
not sent the request to copy the whole file? Presumably long
timeout/errors might be a concern, but
that could happen with ranges too. In any case, if sent the whole
file copy request,
the server file system can figure out the holes and copy more efficiently.

In the case where it is copying local to remote or remote to local -
figuring out whether it is
sparse and optimizing makes a lot of sense - but I didn't think cp did
that (at least the
sections of code I was looking at).



--
Thanks,

Steve