On 09/26/2017 11:58 PM, Hannes Reinecke wrote: > On 09/26/2017 08:24 PM, Bart Van Assche wrote: >> On Tue, 2017-09-26 at 10:22 -0700, Lee Duncan wrote: >>> The SCSI ioctl reset path is smart enough to set the >>> flag tmf_in_progress when a user-requested reset is >>> processed, but it does not wait for IO that is in >>> flight. This can result in lost IOs and hung >>> processes. We should wait for a reasonable amount >>> of time for either the IOs to complete or to fail >>> the request. >> >> Hello Lee, >> >> I'm using this functionality all the time to test how SCSI target code handles >> TMFs while SCSI commands are in progress. So I would regret if the SCSI reset >> ioctl code would be modified such that it waits for outstanding requests. >> Isn't the behavior you described a SCSI LLD bug? Shouldn't such bugs be fixed >> instead of implementing a work-around in the SCSI core? >> > Well, thing is that there is an asymmetry here; originally all SCSI EH > functions were supposed to run with no I/O in flight. > (I've modified that with the asynchronous ABORT TASK TMF, but still). > But when called with sg_reset this is no longer true, we're disallowing > new requests, but do not wait for the in-flight I/O to complete. > And we've had a customer report where calling sg_reset -t on an iSCSI > device caused I/O to become stuck as the in-flight I/O was terminated by > the target reset, but the iSCSI stack never sent a completion for that I/O. > > However, we could also defer this problem until my SCSI EH rework goes > in; that clears up the sg_reset path and might clarify things a bit. > > Cheers, > > Hannes > I will wait and see if the problem still exists there, and address it if it does. Thank you. -- Lee Duncan SUSE Labs
