On Tue, Jun 21, 2022 at 06:48:39PM -0400, Gabriel Krisman Bertazi wrote: > Hi Ming, > > I wrote this against your devel-v3 branch. I'm wondering if you plan to > send a new version of the kernel patch soon? From the latest Yeah, that is on my todo list: https://github.com/ming1/linux/commits/my_for-5.19-ubd-devel_v3 there has lots cleanup & improvement. > discussions, I don't think there were major issues found on review. :) One problem is the driver name, and Christoph thought we have 'arch/um/drivers/ubd*.c'. Not thought of one good candidate yet. > > I hope people don't mind I cc'd linux-block about this userspace code. > Please, let me know if I shouldn't do that. > > -- >8 -- > > The number of sectors read/written is used to verify forward progress of > the request inside the kernel. If we return 0 here, the kernel > understands that as an IO failure (see first check in ubd_complete_rq), > and will reissue the request, causing an infinite loop of unfullfilled > requests. This can be reproduced with: > > ubdsrv/ubd add -t null -n0 -q1 -d1 > dd if=/dev/vda of=/dev/ubdb0 count=1 bs=4k > > The approach minics nullblk, which returns the total IO size. > > Signed-off-by: Gabriel Krisman Bertazi <krisman@xxxxxxxxxxxxx> > --- > tgt_null.c | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/tgt_null.c b/tgt_null.c > index 85636c405f0c..61850a2cd046 100644 > --- a/tgt_null.c > +++ b/tgt_null.c > @@ -20,7 +20,9 @@ static int null_init_tgt(struct ubdsrv_tgt_info *tgt, int type, int argc, > static int null_handle_io_async(struct ubdsrv_queue *q, struct ubd_io *io, > int tag) > { > - ubdsrv_mark_io_done(io, 0); > + const struct ubdsrv_io_desc *iod = ubdsrv_get_iod(q, tag); > + > + ubdsrv_mark_io_done(io, iod->nr_sectors << 9); This issue was actually fixed in master branch, and there are actually lots of change in master: - switch to liburing - switch to c++, and use c++20 coroutine for ->handle_io_async() - all kinds of cleanup I am actually working on ubd-qcow2 with above, but that may take a while and the code isn't posted yet. The motivation is that the whole framework can get verified well enough with one complicated target implementation. Thanks, Ming
