On Mon, Nov 27, 2023 at 02:54:21PM +0200, Matti Vaittinen wrote: > pe 24. marrask. 2023 klo 19.26 Greg Kroah-Hartman > (gregkh@xxxxxxxxxxxxxxxxxxx) kirjoitti: > > > > On Fri, Nov 24, 2023 at 05:32:34PM +0100, Oleksij Rempel wrote: > > > On Fri, Nov 24, 2023 at 03:56:19PM +0000, Greg Kroah-Hartman wrote: > > > > On Fri, Nov 24, 2023 at 03:49:46PM +0000, Mark Brown wrote: > > > > > On Fri, Nov 24, 2023 at 03:27:48PM +0000, Greg Kroah-Hartman wrote: > > > > > > On Fri, Nov 24, 2023 at 03:21:40PM +0000, Mark Brown wrote: > > > > > > > > > > > > This came out of some discussions about trying to handle emergency power > > > > > > > failure notifications. > > > > > > > > > > > I'm sorry, but I don't know what that means. Are you saying that the > > > > > > kernel is now going to try to provide a hard guarantee that some devices > > > > > > are going to be shut down in X number of seconds when asked? If so, why > > > > > > not do this in userspace? > > > > > > > > > > No, it was initially (or when I initially saw it anyway) handling of > > > > > notifications from regulators that they're in trouble and we have some > > > > > small amount of time to do anything we might want to do about it before > > > > > we expire. > > > > > > > > So we are going to guarantee a "time" in which we are going to do > > > > something? Again, if that's required, why not do it in userspace using > > > > a RT kernel? > > > > > > For the HW in question I have only 100ms time before power loss. By > > > doing it over use space some we will have even less time to react. > > > > Why can't userspace react that fast? Why will the kernel be somehow > > faster? Speed should be the same, just get the "power is cut" signal > > and have userspace flush and unmount the disk before power is gone. Why > > can the kernel do this any differently? > > > > > In fact, this is not a new requirement. It exist on different flavors of > > > automotive Linux for about 10 years. Linux in cars should be able to > > > handle voltage drops for example on ignition and so on. The only new thing is > > > the attempt to mainline it. > > > > But your patch is not guaranteeing anything, it's just doing a "I want > > this done before the other devices are handled", that's it. There is no > > chance that 100ms is going to be a requirement, or that some other > > device type is not going to come along and demand to be ahead of your > > device in the list. > > > > So you are going to have a constant fight among device types over the > > years, and people complaining that the kernel is now somehow going to > > guarantee that a device is shutdown in a set amount of time, which > > again, the kernel can not guarantee here. > > > > This might work as a one-off for a specific hardware platform, which is > > odd, but not anything you really should be adding for anyone else to use > > here as your reasoning for it does not reflect what the code does. > > I was (am) interested in knowing how/where the regulator error > notifications are utilized - hence I asked this in ELCE last summer. > Replies indeed mostly pointed to automotive and handling the under > voltage events. > > As to what has changed (I think this was asked in another mail on this > topic) - I understood from the discussions that the demand of running > systems with as low power as possible is even more > important/desirable. Hence, the under-voltage events are more usual > than they were when cars used to be working by burning flammable > liquids :) > > Anyways, what I thought I'd comment on is that the severity of the > regulator error notifications can be given from device-tree. Rationale > behind this is that figuring out whether a certain detected problem is > fatal or not (in embedded systems) should be done by the board > designers, per board. Maybe the understanding which hardware should > react first is also a property of hardware and could come from the > device-tree? Eg, instead of having a "DEVICE_SHUTDOWN_PRIO_STORAGE" > set unconditionally for EMMC, systems could set shutdown priority per > board and per device explicitly using device-tree? Yes, using device tree would be good, but now you have created something that is device-tree-specific and not all the world is device tree :( Also, many devices are finally moving out to non-device-tree busses, like PCI and USB, so how would you handle them in this type of scheme? thanks, greg k-h
