On Thu, Apr 09, 2020 at 11:18:22PM -0700, Christoph Hellwig wrote:
> On Fri, Apr 10, 2020 at 01:53:49AM +0900, Johannes Thumshirn wrote:
> > + if (req_op(rq) == REQ_OP_ZONE_APPEND) {
> > + ret = sd_zbc_prepare_zone_append(cmd, &lba, nr_blocks);
> > + if (ret) {
> > + scsi_free_sgtables(cmd);
> > + return ret;
> > + }
> > + }
>
> So actually.
>
> I've been trying to understand the lifetime of the sgtables. Shouldn't
> we free them in the midayer ->init_command returned BLK_STS_*RESOURCE
> instead? It seems like this just catches one particular error instead
> of the whole category? The end of scsi_queue_rq seem like a particular
> good place, as that also releases the resources for the "hard" errors.
Looking more the situation seems even worse. If scsi_mq_prep_fn
isn't successfull we never seem to free the sgtables, even for fatal
errors. So I think we need a real bug fix here in front of the series.
