On 6/20/22 16:03, Hannes Reinecke wrote: > On 6/20/22 08:28, Christoph Hellwig wrote: >> On Mon, Jun 20, 2022 at 08:24:24AM +0200, Hannes Reinecke wrote: >>> So my idea for SATA is simply _not_ to use reserved tags. >>> Any TMF functions (or the equivalent thereof) should always be sent as >>> non-NCQ commands. And when doing so we're back to QD=1 on SATA anyway, so >>> there _must_ be tags available. Consequently the main reason for having >>> reserved tags (namely to guarantee that tags are available for TMF) doesn't >>> apply here. >> >> At least in the non-elevator case (which includes all passthrough I/O) >> request have tags assigned as soon as they are allocated. So, we >> absolutely can have all tags allocated and then need to do a TMF. > > SATA internals may come to the rescue here; if there's an error all NCQ > commands are aborted. So we'll get at least one command tag back. > As for the command duration limits I'm still waiting for clarification > from Damien if we can reuse tags there. We cannot easily reuse tags as the CDL case is *not* an error. So no queue abort come with it. If we had the queue abort, things would be easy as that would essentially be a regular eh. Simply using scsi_execute_req() does not work since we have no guarantees we can get a tag: all 32 commands being executed may complete with needing sense data, and so scsi_execute_req() would hang waiting for a tag. I could try hacking scsi_execute_req() to reuse a tag instead of allocating a new one... Calling scsi_execute_req() from libata is really ugly though. > > But I do agree it's ugly. > > Cheers, > > Hannes -- Damien Le Moal Western Digital Research
