On Thu, Sep 27, 2018 at 7:01 AM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > > On Thu, Sep 27, 2018 at 06:32:47AM +0000, Avri Altman wrote: > > Also, in this context there is the series in > > https://www.spinics.net/lists/linux-scsi/msg123479.html > > which allows to send UPIUs via a bsg device. > > > > It's not a provisioning series per-se like Evan's and Sayali's. > > It covers the provisioning functionality, > > But also allow to send task management UPIU, and UIC commands, > > Which can be used for testing and validation. > > And as someone having been involved with review of a few different > UFS provisioning bits this is what I think we should be merging. > > Instead of being in a rat race of adding ever new sysfs or configfs > attributes for things that don't matter to normal driver operation > I'd rather have a relatively clean pass through interface and move > policy to userspace. Especially given that there are plenty of > vendor specific commands at these levels as well. There's no policy in my patches (nor Sayali's), nor are there any vendor-specific commands here. The sysfs interface has exposed knobs defined by the UFS specification, which to me seems like the kernel providing a sane abstraction of device functionality. I don't see there being a rat race, as these attributes and the config descriptor are all that's needed to provision a device, and any reasonable future versions of the UFS spec would likely be backwards compatible with respect to attributes and flags. The patches Avri linked to seem fine as well, but I don't see why there's not room for both the "roll your own driver completely in user mode" approach, and the "kernel provides a reasonable abstraction of device functionality" approach to co-exist. We do the same sort of thing for simple buses like I2C for example, where you can both write a kernel driver or do bus transactions directly from user mode. -Evan
