On Thu, Jun 21, 2018 at 04:22:22PM +0800, jianchao.wang wrote:

quoted

Trace, please.  With the latest kernel.  I'm not saying that there
is nothing to fix, but the mode of never completing once timeout
requests as currently done is SCSI is clearly broken.

I didn't find the existing method to simulate this.
So I modified the scsi-debug as following patch as install it as following:
modprobe scsi-debug delay=-1 ndelay=-1
Both 4.17-rc1 and 4.18-rc1 with this patch set could survive from the test.

What tree is this against?  I can't apply it to either current Linus'
tree or 4.17 for that matter.

Also I'm not sure this blk_abort_request call is representative
of the real world.  Drivers do drain their queues before calling
it in general, e.g. take a look at ata_eh_set_pending for the
probably most common user.