On Tue, Jul 31 2018, Brian Norris wrote: > Hi Neil, Boris, > > On Tue, Jul 31, 2018 at 10:12:55PM +0200, Boris Brezillon wrote: >> On Tue, 31 Jul 2018 11:05:11 +1000 >> NeilBrown wrote: >> > On Fri, Jul 27 2018, Boris Brezillon wrote: >> > > On Fri, 27 Jul 2018 11:33:13 -0700 >> > > I'll leave Neil some time to review/test/comment on the patch before >> > > queuing it, but it looks good to me. >> > >> > Thanks. >> > I can confirm that if I apply this patch, my system won't reboot >> > properly (as expected), and if I then add >> > >> > broken-flash-reset; >> > >> > to the jedec,spi-nor device, it starts functioning correctly again. >> > >> > I don't like the pejorative "broken", and it also suggests that a thing >> > used to work, but something happened to break it - this is not >> > accurate. >> > I would prefer something like "reset-not-connected" which is an accurate >> > description of the state of the hardware. > > One reason I didn't specifically say something like "not connected", is > because IIUC it's actually *possible* to have a robust boot sequence > without the RESET# pin -- e.g., if your boot ROM hardcoded a software > reset command (just because it's not really standardized doesn't mean > one can't do it). Yes, if we could change the hardware (ROM is hardware) there are various things we could do to improve reliability. What we want to do in devicetree is to describe the (unchangeable) hardware so that Linux can work with it as well as possible. If I have hardware that doesn't reset the flash on reset, then labeling it doesnt-reset-flash-on-system-reset is perfectly appropriate. Labeling it "broken" is pejorative and unhelpful. > >> > I also think that having a WARN_ON is an over-reaction. Certainly a >> > warning could be appropriate, but just one pr_warn() should be enough. >> > The "problem" is unlikely in practice, and loudly warning people that an >> > asteroid might kill them isn't particularly helpful. >> > >> > I genuinely think that if the system fails to reboot, then Linux is at >> > fault. I accept that changing Linux to be completely robust might be >> > more trouble than it is worth, but I don't accept that it is impossible. > > Did you read my last response on the original thread? In my > understanding, there's always a way to, e.g., b0rk your exception > handlers, etc., such that you cannot guarantee your software fallbacks > will work. Normally, one would rely on a (hardware) watchdog to do your > last resort reset for you, but if said reset cannot also reset your boot > flash, then...you're stuck. > > IOW, it's impossible. I cannot say for certain if I read your last response, but I've read quite a few opinions while researching this and think I have a good handle on the details. I agree that if you want high reliability then you need a properly configured hardware watchdog. Not everyone needs that and not everyone bothers with a watchdog. If you do want a watchdog, you would (obviously?) make sure to buy hardware that supports a watchdog. But if you choose to buy hardware that doesn't have a watchdog, then it isn't "broken", it simply doesn't have a watchdog and so can be expected to freeze if something particularly bad happens. Linux could get almost arbitrarily sophisticated in ensuring that the panic-handling code was fully robust and was stored in write-protected memory, and so be able to reboot cleanly after any panic. There will, of course, be situations where it cannot recover (it might not panic...), but the fact that it needs to reset the flash as part of recovery shouldn't increase the set of such situations noticeably. > > Is that not an accurate description? > >> > But I don't intend to fight either of these battles. >> >> Does that mean you're accepting this change? Brian, any comment on what >> Neil said? >> >> To be honest, I hate being in the middle of this discussion without >> having been involved in the first decision to accept such workarounds. >> I keep thinking that making boards that do not have reset properly >> wired less likely to fail rebooting is a wise decision, but I also >> agree with Brian when he says we should inform people that their design >> is unreliable. >> The main problem I see here, is that adding this prop won't help people >> figuring out what is wrong with their design, it will just help them > > How else would we help someone figure out what's wrong with their > design? My best attempt is to make it quite obvious, as long as they're > using vanilla mainline: if their system hangs on reboot (without this > property), then it's probably a bad design. Is it really our job to help people figure out what's wrong with their designs (unless they ask)? I see it as our job to make Linux work reliably. If a system hangs on reboot, but we can fix reboot so that it doesn't, I think we should. Clearly you disagree. To clearly state my position: 1/ A clean reboot should reboot cleanly, resetting any hardware that might need resetting. 2/ an unclean reboot is never guaranteed (though "best effort" is still a good goal). If you need guaranteed unclean reboots, you need a properly configured hardware watchdog. My hardware doesn't have a properly configured hardware watchdog, and I don't expect it to handle an unclean reboot. I do expect it to handle a clean reboot. I'd rather not be told the hardware is "broken" because it isn't - it simply doesn't have watchdog support (it doesn't have hardware floating point either - that doesn't make it 'broken'). Thanks, NeilBrown > > And if instead, someone stuck in this DT property already, the loud > warning might suggest the reader look at the DT binding doc or code > comments, where I elaborated. > >> workaround the problem when they find out, and it might already be to >> late to fix the HW design. But maybe it's not what we're trying to do >> here. Maybe we just want to warn users that rebooting such boards is a >> risky procedure. > > Brian