Re: scsi error handling thread and REQUEST SENSE

[Bart Van Assche <bvanassche@acm.org>]

On 05/16/14 21:02, Elliott, Robert (Server Storage) wrote:
> The command is still outstanding; data transfers might still occur, 
> and a completion using its tag could still appear at any time. 
> However, the error handler declares that the command is done, 
> so all the buffers are freed and the tag is reused.
> 
> The SCSI error handler needs to escalate this to a reset that 
> ensures that the command is no longer outstanding: ABORT
> TASK (which already didn't work), ABORT TASK SET, LOGICAL 
> UNIT RESET, I_T NEXUS RESET, or hard reset.

If my interpretation of the SCSI mid-layer source code is correct then
even with the patch "improved eh timeout handler" applied the SCSI
mid-layer still guarantees for each SCSI host that at most one
eh_abort_handler() call is active at any given time (since tmf_work_q is
created with max_active = 1) and also that at least one of the eh_*
functions is invoked before the SCSI mid-layer finishes a command. Does
your comment mean that you have found a scenario in which none of the
LLD eh_* callback functions was invoked before the SCSI mid-layer
finished a SCSI command ?

Thanks,

Bart.