On Wed, 04/04 14:23, Stefan Hajnoczi wrote: > On Thu, Mar 29, 2018 at 07:09:06PM +0800, Fam Zheng wrote: > > [Posting a preview RFC for the general idea discussion and internal API > > review. > > Libiscsi support is being worked on in the meantime.] > > > > This series introduces block layer API for copy offloading and makes use of > > it > > in qemu-img convert. > > > > For now we implemented the operation in local file protocol with > > copy_file_range(2). Besides that it's possible to add similar to iscsi, nfs > > and potentially more. > > > > As far as its usage goes, in addition to qemu-img convert, we can emulate > > offloading in scsi-disk (EXTENDED COPY), and do similar to drive-mirror. > > > > The new bdrv_co_map_range can also be an alternative way to implement format > > drivers in the future, once we make block/io.c use it in preadv/pwritev > > paths. > > I posted concerns about the bdrv_co_map_range() interface. It would be > safer to only have a copy_range() interface without exposing how data is > mapped outside the driver where race conditions can occur and the format > driver no longer has full control over file layout.
It's a good point, but I couldn't think of a way to implement copy_range between two format drivers: both of them need to recurse down to their bs->file and what we eventually want is a copy_file_range() on two fds (or an iscsi equivalent): src[qcow2] -> dst[raw] | | v v src[file] -> dst[file] | | v v fd1 -> fd2 copy_file_range Maybe we should add BlockDriver.bdrv_co_map_range_{prepare,commit,abort} and call them from bdrv_co_copy_range(). This way the code path works pretty much the same way to .bdrv_co_preadv/pwritev. Fam