> > On Mon, 2020-12-07 at 19:35 +0100, Greg KH wrote: > > > On Mon, Dec 07, 2020 at 06:26:03PM +0000, Christoph Hellwig wrote: > > > > On Mon, Dec 07, 2020 at 07:23:12PM +0100, Greg KH wrote: > > > > > What "real workload" test can be run on this to help show if it > > > > > is useful or not? These vendors seem to think it helps for some > > > > > reason, otherwise they wouldn't have added it to their silicon :) > > > > > > > > > > Should they run fio? If so, any hints on a config that would be > > > > > good to show any performance increases? > > > > > > > > A real actual workload that matters. Then again that was Martins > > > > request to even justify it. I don't think the broken addressing > > > > that breaks a whole in the SCSI addressing has absolutely not > > > > business being supported in Linux ever. The vendors should have > > > > thought about the design before committing transistors to something > > > > that fundamentally does not make sense. > > > > Actually, that's not the way it works: vendors add commands because > > standards mandate. That's why people who want weird commands go and > > join standard committees. Unfortunately this means that a lot of the > > commands the standard mandates end up not being very useful in > > practice. For instance in SCSI we really only implement a fraction of > > the commands in the standard. > > > > In this case, the industry already tried a very similar approach with > > GEN 1 hybrid drives and it turned into a complete disaster, which is > > why the mode became optional in shingle drives and much better modes, > > which didn't have the huge shared state problem, superseded it. Plus > > truncating the LBA of a READ 16 to 4 bytes is asking for capacity > > problems down the line, so even the actual implementation seems to be > > problematic. > > > > All in all, this looks like a short term fix which will go away when > > the drive capacity improves and thus all the effort changing the driver > > will eventually be wasted. > > "short term" in the embedded world means "this device is stuck with this > chip for the next 8 years", it's not like a storage device you can > replace, so this might be different than the shingle drive mess. Also, > I see many old SoCs still showing up in brand new devices many many > years after they were first introduced, on-chip storage controllers is > something we need to support well if we don't want to see huge > out-of-tree patchsets like UFS traditionally has been lugging around for > many years. > > > > So "time to boot an android system with this enabled and disabled" > > > would be a valid workload, right? I'm guessing that's what the > > > vendors here actually care about, otherwise there is no real stress- > > > test on a UFS system that I know of. > > > > Um, does it? I don't believe even the UFS people have claimed this. > > The problem is that HPB creates a shared state between the driver and > > the device. That shared state has to be populated, which has to happen > > at start of day, so it's entirely unclear if this is a win or a slow > > down for boot. > > Ok, showing that this actually matters is a good rule, Daejun, can you > provide that if you resubmit this patchset? > Sure, I will find out the case which has performance benefit by HPB. Thanks, Daejun