On 05/25/2011 03:24 PM, scameron@xxxxxxxxxxxxxxxxxx wrote: >> Le 25 mai 2011 06:50, <scameron@xxxxxxxxxxxxxxxxxx> a écrit : >>> Or, maybe I should be reporting the CMD_ABORTED case back as >>> >>> cmd->result = DID_SOFT_ERROR << 16; >> >> look at scsi_error.c:scsi_decide_disposition(). DID_SOFT_ERROR >> triggers retry handling while DID_ABORT does not (at least in the tree >> that i'm looking at). >> > > Yes this is what I was looking at, and part of what is confusing me. > This makes me think that DID_ABORT is the wrong thing to do, because > it would seem that there is no retry, and yet, dd does not complain > that one of the reads initiated on it's behalf has been aborted. > Does that mean it copied bogus data (uninitialized memory)? Not sure, > but if so, that would be bad. (Easy enough to test this case.) scsi_decide_disposition is not used in the abort path. In the abort path commands go through scsi_eh_flush_done_q. In there you can see some basic handling. We check if we can failfast the cmd (scsi_noretry_cmd) and if there are enough retries. If it is not fastfailbale and there are enough retries we retry the IO by calling scsi_queue_insert. If there are not enough retries or we want to fastfail it then scsi_finish_command is called like is done in the normal completion path from (scsi_softirq_done -> scsi_decide_disposition -> (look at return value) -> scsi_finish_command. >From scsi_finish_command the cmd is then handled like in the normal path where something like the ULD done function is called then scsi_io_completion is. So as you can see from scsi_eh_flush_done_q, the scsi layer does not really look at what you return and do anything with it. There is just the basic check for if the IO is fastfailable. When I added that code, it looked like most drivers returned a wide range of values and in some cases did not return anything (if a driver does not set anything for the result then there is that check in scsi_eh_flush_done_q that sets the DRIVER_TIMEOUT bit). I do not know what the policy is. I have been converting drivers to return something useful when they want to control if IO is fastfailed in this path. -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
