Mike Christie wrote:

quoted

I tested this but it does not quite work: there are now two retries in 
the same test that previously gave six retries. Unfortunately I don't 
just now have time to study the problem further.

ok I will try to look into it.

Oh I bet it is from the change that James B noticed. With scsi_do_req() 
we passed in the sr_done function, but with scsi_execute* we pass in the 
requests end_io function. So with scsi_do_req, when we complete a 
command in scsi_finish_command()  (where it calls cmd->done) we would 
call the done function passed into scsi_do_req(), but with scsi_execute* 
we call scsi_io_completion.

Maybe it could be solved by having st implement a cmd->done function 
like sd.c that gets set in the init_command callout. Since all the 
commands will be coming down as REQ_BLOCK_PC commands, st can set the 
scsi_cmnd->done function and decide what to do there. Or maybe some of 
the scsi_io_completion logic needs to be fixed. What was the command 
that is causing the problems?