On 05/22/2017 09:54 AM, hch@xxxxxx wrote: > On Mon, May 22, 2017 at 08:06:46AM +0200, Hannes Reinecke wrote: >> Problem is that quite some LLDDs require a hardware tag for sending >> TMFs, and currently the re-use the tag from the passed in scmd for that. >> So first we need to move those to a sane interface, and having them >> requesting a new tag for TMFs. > > Some do indeed. But these seems to be the exception and not the rule. > Still need to be handled, though ... >> Should be made easier with Christophs rework, but we still don't have >> any defined way how TMFs should request their tag; should they use >> private tags? Should they use 'normal' tags? Should the driver implement >> their own tag pool, seeing that these commands will never ever making >> use of the associated request? > > I think the right way to go is that every driver that needs tags for > TMFs should use private tags internally without the core knowing > about it. > That was my idea, too. >> Hence I'm looking into implementing a REQ_RESET block request operation, >> which then could be used to facilitate all of this (the request would be >> allocated from the private tag pool if present). > > That seems to be overkill to me for the few drivers. And I suspect > most of them would be better off now even using blk-mq private tags > (which we'd have to implement for the legacy path first or just > kill it off) but just not expose a tag per host to the scsi and block > layers and set that aside. > IE not using blk-mq private tags for EH? Hmm. But then we'd need a SCSI-internal mechanism to get one of them. I really would want to avoid having each driver to implement it's own mechanism on how to get a TMF tag; that sort of thing only leads to copy-and-paste errors. Ok; will be looking into it. >> It would also neatly solve the scsi_ioctl_reset() problem, as we then >> could just issue a REQ_RESET and would avoid having to call into the >> eh_* function directly. > > I don't think calling the eh_* methods is a problem per see. It's just > their stupid calling conventions.. > I know. I've been working on a patchset to move at least eh_host_reset() to take the scsi_host as argument. But even that requires some preliminary patches to some LLDDs :-( Right. Dusting off those patches, then. Cheers, Hannes -- Dr. Hannes Reinecke zSeries & Storage hare@xxxxxxxx +49 911 74053 688 SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg GF: F. Imendörffer, J. Smithard, D. Upmanyu, G. Norton HRB 21284 (AG Nürnberg)
