Having an issue with vioscsi in GCE. vioscsi decrements the target id in
commands it sends to virtio. This means that the upper layers notion of the
target id is off by 1. This is especially problematic when presented a lun
on the last target of a bus. It never gets discovered because the scsipi
probe code sends down an xs with the last target number, which vioscsi then
decrements in the command it sends along. The probe code never moves on to
the next target (since it thinks its probed all the targets on the bus) and
the lun is never discovered. I hit this in GCE for local disks because they
create a new bus with a single target for each local disk.
Here's a patch that fixes the issue for me:
@@ -279,12 +279,12 @@ vioscsi_scsipi_request(struct scsipi_channel *chan,
scsipi_adapter_req_t
goto stuffup;
}
req->lun[0] = 1;
- req->lun[1] = periph->periph_target - 1;
+ req->lun[1] = periph->periph_target;
req->lun[2] = 0x40 | (periph->periph_lun >> 8);
req->lun[3] = periph->periph_lun;
memset(req->lun + 4, 0, 4);
DPRINTF(("%s: command for %u:%u at slot %d\n", __func__,
- periph->periph_target - 1, periph->periph_lun, slot));
+ periph->periph_target, periph->periph_lun, slot));
if ((size_t)xs->cmdlen > sizeof(req->cdb)) {
DPRINTF(("%s: bad cmdlen %zu > %zu\n", __func__,
Any reason for decrementing periph_target in the first place?
