On 20/06/2017 at 22:15:36 +0100, Russell King - ARM Linux wrote:
On Tue, Jun 20, 2017 at 05:07:46PM +0200, Benjamin Gaignard wrote:
2017-06-20 15:48 GMT+02:00 Alexandre Belloni alexandre.belloni@free-electrons.com:
Yes, that's argument against changing rtc _drivers_ for hardware that can not do better than 32bit. For generic code (such as 44/51 sysfs, 51/51 suspend test), the change still makes sense.
What I had in mind when writing those patches was to remove the limitations coming from those functions usage, even more since they been marked has deprecated.
I'd say that they should not be marked as deprecated. They're entirely appropriate for use with hardware that only supports a 32-bit representation of time.
It's entirely reasonable to fix the ones that use other representations that exceed that, but for those which do not, we need to keep using the 32-bit versions. Doing so actually gives us _more_ flexibility in the future.
Consider that at the moment, we define the 32-bit RTC representation to start at a well known epoch. We _could_ decide that when it wraps to 0x80000000 seconds, we'll define the lower 0x40000000 seconds to mean dates in the future - and keep rolling that forward each time we cross another 0x40000000 seconds. Unless someone invents a real time machine, we shouldn't need to set a modern RTC back to 1970.
I agree with that but not the android guys. They seem to mandate an RTC that can store time from 01/01/1970. I don't know much more than that because they never cared to explain why that was actually necessary (apart from a laconic "this will result in a bad user experience")
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly. My opinion is that as long as userspace is not ready to handle those dates, it doesn't really matter because it is quite unlikely that anything will be able to continue running anyway.
On Tue, 20 Jun 2017, Alexandre Belloni wrote:
On 20/06/2017 at 22:15:36 +0100, Russell King - ARM Linux wrote:
On Tue, Jun 20, 2017 at 05:07:46PM +0200, Benjamin Gaignard wrote:
2017-06-20 15:48 GMT+02:00 Alexandre Belloni alexandre.belloni@free-electrons.com:
Yes, that's argument against changing rtc _drivers_ for hardware that can not do better than 32bit. For generic code (such as 44/51 sysfs, 51/51 suspend test), the change still makes sense.
What I had in mind when writing those patches was to remove the limitations coming from those functions usage, even more since they been marked has deprecated.
I'd say that they should not be marked as deprecated. They're entirely appropriate for use with hardware that only supports a 32-bit representation of time.
It's entirely reasonable to fix the ones that use other representations that exceed that, but for those which do not, we need to keep using the 32-bit versions. Doing so actually gives us _more_ flexibility in the future.
Consider that at the moment, we define the 32-bit RTC representation to start at a well known epoch. We _could_ decide that when it wraps to 0x80000000 seconds, we'll define the lower 0x40000000 seconds to mean dates in the future - and keep rolling that forward each time we cross another 0x40000000 seconds. Unless someone invents a real time machine, we shouldn't need to set a modern RTC back to 1970.
I agree with that but not the android guys. They seem to mandate an RTC that can store time from 01/01/1970. I don't know much more than that because they never cared to explain why that was actually necessary (apart from a laconic "this will result in a bad user experience")
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly.
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it?
2) Deal with creative vendors who have their own idea about the 'start of the epoch' 3) Add the information of wraparound time to the rtc device which needs to be filled in for each device. That way the rtc_*** accessor functions can deal with them whether they wrap in 2038 or 2100 or whatever.
#3 is the simplest problem of them :)
My opinion is that as long as userspace is not ready to handle those dates, it doesn't really matter because it is quite unlikely that anything will be able to continue running anyway.
That's a different story. Making the kernel y2038 ready in general is a good thing. Whether userspace will be ready by then or not is completely irrelevant.
Thanks,
tglx
On Wed, Jun 21, 2017 at 12:00:30AM +0200, Thomas Gleixner wrote:
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it? 2) Deal with creative vendors who have their own idea about the 'start of the epoch'3) Add the information of wraparound time to the rtc device which needs to be filled in for each device. That way the rtc_*** accessor functions can deal with them whether they wrap in 2038 or 2100 or whatever.
#3 is the simplest problem of them :)
Well, if there's additional non-volatile storage, you can store additional information in there, but you still need the RTC subsystem to be aware that the hardware is only 32-bit capable.
You'd also still need something along the lines I detailed (redefining what dates the past 32-bit values indicate) to cope with the RTC being set backwards after the machine thinks (possibly incorrectly) that the date has jumped forwards.
Hi!
I agree with that but not the android guys. They seem to mandate an RTC that can store time from 01/01/1970. I don't know much more than that because they never cared to explain why that was actually necessary (apart from a laconic "this will result in a bad user experience")
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly.
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it?
I'd not do offset. Instead, I'd select a threshold (perhaps year of release of given kernel?) and
if (rtc_time < year_of_release_of_kernel) rtc_time += 0x100000000;
Ok, we'll have to move away from "rtc_time == 0 indicates zero", as seen in some drivers.
2) Deal with creative vendors who have their own idea about the 'start of the epoch'
If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Pavel
On 21/06/2017 at 09:51:52 +0200, Pavel Machek wrote:
Hi!
I agree with that but not the android guys. They seem to mandate an RTC that can store time from 01/01/1970. I don't know much more than that because they never cared to explain why that was actually necessary (apart from a laconic "this will result in a bad user experience")
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly.
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it?I'd not do offset. Instead, I'd select a threshold (perhaps year of release of given kernel?) and
if (rtc_time < year_of_release_of_kernel) rtc_time += 0x100000000;
Ok, we'll have to move away from "rtc_time == 0 indicates zero", as seen in some drivers.
2) Deal with creative vendors who have their own idea about the 'start of the epoch'If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Hi!
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly.
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it?I'd not do offset. Instead, I'd select a threshold (perhaps year of release of given kernel?) and
if (rtc_time < year_of_release_of_kernel) rtc_time += 0x100000000;
Ok, we'll have to move away from "rtc_time == 0 indicates zero", as seen in some drivers.
2) Deal with creative vendors who have their own idea about the 'start of the epoch'If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
...but still better than board stuck in the past, no?
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Well, you can still convert time from rtc to unix time, then do adjustment there.
Anyway, I guess it would be cool for rtc drivers to annotate what limits underlying storage has to the common code, so that we can do fixups once per class, not once per driver.
On 21/06/2017 at 08:34:43 +0200, Pavel Machek wrote:
Hi!
I think tglx had a plan for offsetting the time at some point so 32-bit platform can pass 2038 properly.
Yes, but there are still quite some issues to solve there:
1) How do you tell the system that it should apply the offset in the first place, i.e at boot time before NTP or any other mechanism can correct it?I'd not do offset. Instead, I'd select a threshold (perhaps year of release of given kernel?) and
if (rtc_time < year_of_release_of_kernel) rtc_time += 0x100000000;
Ok, we'll have to move away from "rtc_time == 0 indicates zero", as seen in some drivers.
2) Deal with creative vendors who have their own idea about the 'start of the epoch'If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
...but still better than board stuck in the past, no?
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Well, you can still convert time from rtc to unix time, then do adjustment there.
You can only if your machine is running when that happens. If that is not the case, then you lost and your time is not correct anymore.
There is currently one rtc doing that kind of trick but it is used as a simple time counter from the beginning. Transitioning is the difficult part.
Anyway, I guess it would be cool for rtc drivers to annotate what limits underlying storage has to the common code, so that we can do fixups once per class, not once per driver.
Yes, I'm in the middle of the whole rework that allows that.
I don't understand the sudden urgency of fixing that and the amount of bikeshedding, seeing that the closest cutoff date is actually 31st of december 2069 in the rtc subsystem and that anyway the current 32bit userspace will explode in february 2038.
My plan from the beginning was to have something for the next stable. I know nobody can read my mind but again, I don't think there is currently any urgency to change anything.
Hi!
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
...but still better than board stuck in the past, no?
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Well, you can still convert time from rtc to unix time, then do adjustment there.
You can only if your machine is running when that happens. If that is not the case, then you lost and your time is not correct anymore.
I don't see why that should be a case... as long as you know what RTC does in event of overflow, and it is not something completely crazy.
Anyway, I guess it would be cool for rtc drivers to annotate what limits underlying storage has to the common code, so that we can do fixups once per class, not once per driver.
Yes, I'm in the middle of the whole rework that allows that.
I don't understand the sudden urgency of fixing that and the amount of bikeshedding, seeing that the closest cutoff date is actually 31st of december 2069 in the rtc subsystem and that anyway the current 32bit userspace will explode in february 2038.
My plan from the beginning was to have something for the next stable. I know nobody can read my mind but again, I don't think there is currently any urgency to change anything.
Yes, mind reading is a problem. I can only read minds of ferungulates, and only if they are physicaly near me :-).
Best regards, Pavel
On Wed, Jun 21, 2017 at 10:39:07AM +0200, Alexandre Belloni wrote:
On 21/06/2017 at 09:51:52 +0200, Pavel Machek wrote:
If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
... and that basically means you can't use hardware that stores RTC time as a 32-bit number of seconds past 2106.
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Not for the case being discussed. A 32-bit counter of seconds knows nothing about leap years - all that is handled by the conversion functions.
On 21/06/2017 at 10:19:49 +0100, Russell King - ARM Linux wrote:
On Wed, Jun 21, 2017 at 10:39:07AM +0200, Alexandre Belloni wrote:
On 21/06/2017 at 09:51:52 +0200, Pavel Machek wrote:
If someone uses different threshold, well, there will be confusion. But only for users that have their rtc set to the past, which is quite unusual.
Or not, having an RTC set in the past is actually quite common. I'd find it weird to have a new device boot and be set to a date in the future.
... and that basically means you can't use hardware that stores RTC time as a 32-bit number of seconds past 2106.
And I guess it will not matter much for us anyway ;)
Also note that the threshold or offset thing may seem like a good idea but fails with many RTCs because of how they handle leap years.
Not for the case being discussed. A 32-bit counter of seconds knows nothing about leap years - all that is handled by the conversion functions.
Well, the patch series touches some RTCs that are not using 32 bit counter so I though I might as well raise the issue now.
linaro-kernel@lists.linaro.org
-
Alexandre Belloni -
Pavel Machek -
Russell King - ARM Linux -
Thomas Gleixner