Il 11/02/2013 15:18, Stefan Priebe - Profihost AG ha scritto:
>> > Some trace that a request was actually cancelled, but I think I
>> > believe
> Ah but that must be in guest not on host right? How to grab that from
> client when it is crashing?
Serial console could have something like "sda: aborting command". It is
actually interesting to see what is causing commands to be aborted (typically a
timeout, but what causes the timeout? :).
>> > that. This seems to be the same issue as commits
>> > 1bd075f29ea6d11853475c7c42734595720c3ac6 (iSCSI) and
>> > 473c7f0255920bcaf37411990a3725898772817f (rbd), where the "cancelled"
>> > callback is called before the "complete" callback.
> If there is the same code in virtio-scsi it might be.
No, virtio-scsi is relying on the backends (including scsi-disk)
doing it correctly. The RBD code looks okay, so it's still my
fault :) but not virtio-scsi's.
I think this happens when a request is split into multiple parts,
and one of them is canceled. Then the next part is fired, but
virtio-scsi's cancellation callbacks have fired already.
You can test this patch:
diff --git a/hw/scsi/scsi-disk.c b/hw/scsi/scsi-disk.c
index 07220e4..1d8289c 100644
--- a/hw/scsi/scsi-disk.c
+++ b/hw/scsi/scsi-disk.c
@@ -221,6 +221,10 @@ static void scsi_write_do_fua(SCSIDiskReq *r)
{
SCSIDiskState *s = DO_UPCAST(SCSIDiskState, qdev, r->req.dev);
+ if (r->req.io_canceled) {
+ return;
+ }
+
if (scsi_is_cmd_fua(&r->req.cmd)) {
bdrv_acct_start(s->qdev.conf.bs, &r->acct, 0, BDRV_ACCT_FLUSH);
r->req.aiocb = bdrv_aio_flush(s->qdev.conf.bs, scsi_aio_complete, r);
@@ -352,6 +356,10 @@ static void scsi_read_data(SCSIRequest *req)
/* No data transfer may already be in progress */
assert(r->req.aiocb == NULL);
+ if (r->req.io_canceled) {
+ return;
+ }
+
/* The request is used as the AIO opaque value, so add a ref. */
scsi_req_ref(&r->req);
if (r->req.cmd.mode == SCSI_XFER_TO_DEV) {
@@ -455,6 +463,10 @@ static void scsi_write_data(SCSIRequest *req)
/* No data transfer may already be in progress */
assert(r->req.aiocb == NULL);
+ if (r->req.io_canceled) {
+ return;
+ }
+
/* The request is used as the AIO opaque value, so add a ref. */
scsi_req_ref(&r->req);
if (r->req.cmd.mode != SCSI_XFER_TO_DEV) {
Paolo
