GTOD: persistent clock support, core
Quoted 4 lines
Implement generic timekeeping suspend/resume accounting by introducing the
read_persistent_clock()interface. This is an arch specific function that returns the seconds since the epoch using the arch defined battery backed clock.
I remain unclear what's expected to happen when there IS no such architcture-defined clock ... but where the system itself still has one, e.g. a board may access one through I2C or SPI once IRQs are working normally.
You'll recall that I had pointed out that the drivers/rtc framework provides CONFIG_RTC_HCTOSYS, which already unifies quite a lot of the "persistent" clocks in the way you described above, but without that nasty requirement of working without IRQs enabled.
Quoted 8 lines
+/**
+ *read_persistent_clock- Return time in seconds from the persistent clock.
+ *
+ * Weak dummy function for arches that do not yet support it.
+ * Returns seconds from epoch using the battery backed persistent clock.
+ * Returns zero if unsupported.
+ *
+ * XXX - Do be sure to remove it once all arches implement it.
But not all architectures **CAN** support this notion ...
Quoted 4 lines
+ */
(quoted diff) +3
-0+unsigned long __attribute__((weak)) read_persistent_clock(void) +{ + return 0;
Review
Quoted 10 lines
(quoted diff) +3
-0+} + /* * timekeeping_init - Initializes the clocksource and common timekeeping values */ void __init timekeeping_init(void) { unsigned long flags; + unsigned long sec = read_persistent_clock();
... and timekeeping_init() is called before I2C or SPI could be used, since IRQs aren't enabled yet and accessing those busses can't be done in general without IRQs enabled.
Review static int timekeeping_suspended; · line 804
Quoted 5 lines
(quoted diff) +1
-0@@ -774,13 +801,23 @@ static int timekeeping_suspended; static int timekeeping_resume(struct sys_device *dev) { unsigned long flags; + unsigned long now = read_persistent_clock();
Again: sys_device resume() is called with IRQs disabled, which prevents access to many systems' persistent clocks. In fact, after posting this example patch
http://marc.theaimsgroup.com/?l=linux-kernel&m=115600629813751&w=2
I never heard anything more from you on this issue. Given this particular patch (in $SUBJECT) should I assume you're going to just ignore the issues whereby RTCs ("persistent clocks") can't always meet the no-IRQs-needed assumptions being made here? Or address isssues like using pointer-to-function instead of using linker tricks?
http://marc.theaimsgroup.com/?l=linux-kernel&m=115600629825461&w=2
Those class suspend/resume hooks are now merged to kernel.org, by the way, so that example patch is now pretty much deployable.
- Dave
Quoted 9 lines · David Brownell
Quoted 4 lines
Implement generic timekeeping suspend/resume accounting by introducing the
read_persistent_clock()interface. This is an arch specific function that returns the seconds since the epoch using the arch defined battery backed clock.I remain unclear what's expected to happen when there IS no such architcture-defined clock ... but where the system itself still has one, e.g. a board may access one through I2C or SPI once IRQs are working normally.
Yea. First, let me apologize for falling off the last thread. I got busy with other things and the discussion withered. This patch has been re-raised because Thomas finally tripped over the "theoretical" resume time ordering bug that I was concerned about.
So, while my first announcement with the patch was something of the effect of "Trying to unify the cmos/RTC/whatever interface", due to our discussion I'm scaling back that goal.
Instead the purpose of this is just a continuation of the generic timekeeping work. Moving the *existing* arch specific resume timekeeping code into generic code, so we don't get crazy resume ordering issues splitting the issue of who sets what when between the generic and arch specific time resume functions.
On arches that have the constraints you list above, a dummy read_persistent_clock() that returns zero can be implemented, with a comment about why and the RTC_HCTOSYS bits can be used, with no change in behavior from what we have now.
Quoted 15 lines
You'll recall that I had pointed out that the
drivers/rtcframework providesCONFIG_RTC_HCTOSYS, which already unifies quite a lot of the "persistent" clocks in the way you described above, but without that nasty requirement of working without IRQs enabled.Quoted 8 lines
+/**
+ *read_persistent_clock- Return time in seconds from the persistent clock.
+ *
+ * Weak dummy function for arches that do not yet support it.
+ * Returns seconds from epoch using the battery backed persistent clock.
+ * Returns zero if unsupported.
+ *
+ * XXX - Do be sure to remove it once all arches implement it.
Quoted 1 line
But not all architectures **CAN** support this notion ...
That's ok and is the reason why we have a unsupported return option.
This patch just gives us a path to consolidate what the majority of arches do currently.
Quoted 11 lines
Quoted 7 lines
/*
*timekeeping_init- Initializes the clocksource and common timekeeping values */
void__inittimekeeping_init(void){ unsigned long flags;+ unsigned long sec =
read_persistent_clock();... and
timekeeping_init()is called before I2C or SPI could be used, since IRQs aren't enabled yet and accessing those busses can't be done in general without IRQs enabled.
Again, that's fine. If the read_persistent_clock is not supported, xtime will still be zero and can be set later via other methods.
Quoted 11 lines
Review static int timekeeping_suspended; · line 804
Quoted 5 lines
(quoted diff) +1
-0@@ -774,13 +801,23 @@ static int timekeeping_suspended; static int timekeeping_resume(struct sys_device *dev) { unsigned long flags; + unsigned long now = read_persistent_clock();Again:
sys_deviceresume()is called with IRQs disabled, which prevents access to many systems' persistent clocks. In fact, after posting this example patchhttp://marc.theaimsgroup.com/?l=linux-kernel&m=115600629813751&w=2
Ok, correct me if I'm wrong, though: The only catch, if I understand correctly, is that it requires the system in question to have a proper RTC driver, which doesn't cover 100% of the arch/platforms supported. No?
I don't really have an issue w/ the RTC code above, however it does not address the current suspend/resume code in the majority of arches. I don't know if we're actually in that much conflict here, since I'm trying to remove the arch specific suspend/resume timekeeping changes, and (to my understanding) so are you.
We just have a difference in where we're trying to re-add the code. I'm moving the current code into the generic tod path, and you're moving it into the RTC driver. Both efforts are consolidating code, so even with the minor duplication we have less code then before. So I'm sure as we whittle away at this we can find a way to meet in the middle. I think this patch moves us forward in that direction.
Quoted 6 lines
I never heard anything more from you on this issue. Given this particular patch (in $SUBJECT) should I assume you're going to just ignore the issues whereby RTCs ("persistent clocks") can't always meet the no-IRQs-needed assumptions being made here? Or address isssues like using pointer-to-function instead of using linker tricks?
In my head I see it like this:
Currently here is how the timekeeping resume code breaks down:
1) timeofday_resume: reset clocksource, NTP
2) arch time resume: [set xtime]
3) RTC resume: [set xtime]
Where the [set xtime]s depend on platform config
I'm trying to just move us to:
1) timeofday_resume: reset clocksource, NTP, [set xtime]
2) RTC resume: [set xtime]
Again, where the [set xtime]s depend on platform config
Then as the RTC drivers gain coverage, maybe we can start cutting timeofday resume down and have something like:
1) timeofday_resume: reset clocksource, NTP
2) RTC resume: set xtime
Does this sound like a way forward?
thanks
-john
Quoted 5 lines · john stultz
Yea. First, let me apologize for falling off the last thread. I got busy with other things and the discussion withered. This patch has been re-raised because Thomas finally tripped over the "theoretical" resume time ordering bug that I was concerned about.
I recall losing track of what that issue was, and thinking a new explanation was needed ... no such issues were described in this patch or its summary.
Quoted 7 lines
So, while my first announcement with the patch was something of the effect of "Trying to unify the cmos/RTC/whatever interface", due to our discussion I'm scaling back that goal.
Instead the purpose of this is just a continuation of the generic timekeeping work. Moving the *existing* arch specific resume timekeeping code into generic code,
... which wasn't what that one source code comment implied; quite the opposite, it assumed all architectures _could_ work that way.
Quoted 4 lines
On arches that have the constraints you list above, a dummy
read_persistent_clock()that returns zero can be implemented, with a comment about why and theRTC_HCTOSYSbits can be used, with no change in behavior from what we have now.
Better IMO to use a more traditional run-time function call style hook, not the link time magic you're now using. Point is, this is most accurately a _board_ option not an architectural one.
Even systems that _could_ provide some access to a persistent clock without relying on IRQs might be designed not to do so. For example, RTCs integrated into system-on-chip processors can be less power-efficient than discrete ones, so that when power-efficient design is essential, that integrated RTC may not be a viable design option.
And since we expect Linux kernels to be able to run on more than one board, that includes configurations where one of the boards uses the integrated register-based RTC, and another uses a discrete one with serial interface. But your choice of link-time binding precludes one kernel handling both boards correctly...
Quoted 9 lines
Quoted 9 lines
Quoted 8 lines
+/**
+ *read_persistent_clock- Return time in seconds from the persistent clock.
+ *
+ * Weak dummy function for arches that do not yet support it.
+ * Returns seconds from epoch using the battery backed persistent clock.
+ * Returns zero if unsupported.
+ *
+ * XXX - Do be sure to remove it once all arches implement it.
Quoted 6 lines
Quoted 1 line
But not all architectures **CAN** support this notion ...
That's ok and is the reason why we have a unsupported return option.
This patch just gives us a path to consolidate what the majority of arches do currently.
That XXX comment disagrees ... it assumes all architectures can work that way ...
And I'm not sure about "majority"; depends how you count. Certainly the number of embedded boards that could work differently is large and growing ... and there are a lot more embedded computers in the world than desktops, laptops, or servers.
Quoted 10 lines
Quoted 10 lines
Review static int timekeeping_suspended; · line 804
Quoted 5 lines
(quoted diff) +1
-0@@ -774,13 +801,23 @@ static int timekeeping_suspended; static int timekeeping_resume(struct sys_device *dev) { unsigned long flags; + unsigned long now = read_persistent_clock();Again:
sys_deviceresume()is called with IRQs disabled, which prevents access to many systems' persistent clocks. In fact, after posting this example patch
Quoted 7 lines
Ok, correct me if I'm wrong, though: The only catch, if I understand correctly, is that it requires the system in question to have a proper RTC driver, which doesn't cover 100% of the
arch/platformssupported. No?
It's not a "catch" for the systems which have an RTC class driver, and your approach doesn't cover 100% either.
The key difference between the two approaches is that yours **CANNOT** handle cases like an I2C based RTC (call them "message based RTCs"), where the RTC class drivers **CAN** handle "register based RTCs".
Quoted 5 lines
I don't really have an issue w/ the RTC code above, however it does not address the current suspend/resume code in the majority of arches. I don't know if we're actually in that much conflict here, since I'm trying to remove the arch specific suspend/resume timekeeping changes, and (to my understanding) so are you.
It's a question of scope, I guess. You're limiting your solution to boards that don't require message based RTCs. I'm thinking that's not really sufficient ... and it bothers me to see something be pitched as a basic architectural feature of the clock/time framework when it's so clearly not sufficient.
Quoted 6 lines
We just have a difference in where we're trying to re-add the code. I'm moving the current code into the generic tod path, and you're moving it into the RTC driver. Both efforts are consolidating code, so even with the minor duplication we have less code then before. So I'm sure as we whittle away at this we can find a way to meet in the middle. I think this patch moves us forward in that direction.
So long as this GTOD patch is updated to reflect the reality that not all _boards_ can (or will) support that style of RTC ... cleanup is good, but those code comments suggest deeper assumptions.
Quoted 7 lines
Quoted 6 lines
I never heard anything more from you on this issue. Given this particular patch (in $SUBJECT) should I assume you're going to just ignore the issues whereby RTCs ("persistent clocks") can't always meet the no-IRQs-needed assumptions being made here? Or address isssues like using pointer-to-function instead of using linker tricks?
Quoted 8 lines
In my head I see it like this:
Currently here is how the timekeeping resume code breaks down:
1)timeofday_resume: reset clocksource, NTP
2) arch time resume: [set xtime]
3) RTC resume: [set xtime]Where the [set xtime]s depend on platform config
That is, three different ways to do it.
Quoted 5 lines
I'm trying to just move us to:
1)timeofday_resume: reset clocksource, NTP, [set xtime]
2) RTC resume: [set xtime]Again, where the [set xtime]s depend on platform config
Getting rid of one is useful, yes. But why not get rid of two?
Quoted 6 lines
Then as the RTC drivers gain coverage, maybe we can start cutting timeofday resume down and have something like:
1)timeofday_resume: reset clocksource, NTP
2) RTC resume: set xtimeDoes this sound like a way forward?
I'm not sure I see how the last two groups are very different; are you saying it would get rid of these message-unfriendly read_persistent_clock() calls?
- Dave
Quoted 9 lines · David Brownell
Quoted 5 lines · john stultz
Yea. First, let me apologize for falling off the last thread. I got busy with other things and the discussion withered. This patch has been re-raised because Thomas finally tripped over the "theoretical" resume time ordering bug that I was concerned about.
I recall losing track of what that issue was, and thinking a new explanation was needed ... no such issues were described in this patch or its summary.
Right. So briefly, the resume ordering issue is this: Currently resuming the timekeeping code is split across both the generic code and the arch specific code. It wasn't always this way, my earlier GTOD patches had a similar change, but in my reworking and cutting down of the code to get it acceptable for mainline I dropped it.
The assumption I made, was that the generic timekeeping_resume would be called before the arch specific time resume code runs. And by luck it does, however I realized shortly afterward that there isn't a strict ordering to the resume functions and it could happen the other way. With the dynticks patch, Thomas finally tripped over the issue, and so I got this patch going again.
Quoted 10 lines
Quoted 7 lines
So, while my first announcement with the patch was something of the effect of "Trying to unify the cmos/RTC/whatever interface", due to our discussion I'm scaling back that goal.
Instead the purpose of this is just a continuation of the generic timekeeping work. Moving the *existing* arch specific resume timekeeping code into generic code,
... which wasn't what that one source code comment implied; quite the opposite, it assumed all architectures _could_ work that way.
Well to me the comment meant that all arches could implement the function which would return the persistent time or zero(unsupported). And I guess I was thinking the arch implementation it would include comment at to why it returns zero.
But sure, I can understand its confusing and will re-work the wording to be more clear.
Quoted 16 lines
Quoted 4 lines
On arches that have the constraints you list above, a dummy
read_persistent_clock()that returns zero can be implemented, with a comment about why and theRTC_HCTOSYSbits can be used, with no change in behavior from what we have now.Better IMO to use a more traditional run-time function call style hook, not the link time magic you're now using. Point is, this is most accurately a _board_ option not an architectural one.
Even systems that _could_ provide some access to a persistent clock without relying on IRQs might be designed not to do so. For example, RTCs integrated into system-on-chip processors can be less power-efficient than discrete ones, so that when power-efficient design is essential, that integrated RTC may not be a viable design option.
Quoted 6 lines
And since we expect Linux kernels to be able to run on more than one board, that includes configurations where one of the boards uses the integrated register-based RTC, and another uses a discrete one with serial interface. But your choice of link-time binding precludes one kernel handling both boards correctly...
Sure, using a function pointer is easy enough to change.
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Again:
sys_deviceresume()is called with IRQs disabled, which prevents access to many systems' persistent clocks. In fact, after posting this example patchhttp://marc.theaimsgroup.com/?l=linux-kernel&m=115600629813751&w=2
Quoted 7 lines
Quoted 4 lines
Ok, correct me if I'm wrong, though: The only catch, if I understand correctly, is that it requires the system in question to have a proper RTC driver, which doesn't cover 100% of the
arch/platformssupported. No?It's not a "catch" for the systems which have an RTC class driver, and your approach doesn't cover 100% either.
True, it doesn't provide a solution for 100% of the systems. But that's not what I'm shooting for. I'm trying to fix a real bug (the resume ordering issue), and reduce the arch specific logic.
Quoted 3 lines
The key difference between the two approaches is that yours **CANNOT** handle cases like an I2C based RTC (call them "message based RTCs"), where the RTC class drivers **CAN** handle "register based RTCs".
Sure. As a long term solution, the RTC method has promise. The patch in question does not prohibit the RTC work from getting there.
It simply resolves a possible resume ordering bug and allows logic that already exists in alpha, cris, frv, h8300, i386, ia64, m68k, m68knommu, mips, parisc, powerpc, ppc, s390...
[deep inhale]
sh, sh64, sparc64, um, v850, x86_64, and xtensa [whew!] to be consolidated in the generic path.
And the above doesn't challenge your point about there being more embedded systems then desktops. This patch isn't about one arch or platform being more important or more common then another, its just a code duplication and maintenance issue.
Quoted 11 lines
Quoted 5 lines
I don't really have an issue w/ the RTC code above, however it does not address the current suspend/resume code in the majority of arches. I don't know if we're actually in that much conflict here, since I'm trying to remove the arch specific suspend/resume timekeeping changes, and (to my understanding) so are you.
It's a question of scope, I guess. You're limiting your solution to boards that don't require message based RTCs. I'm thinking that's not really sufficient ... and it bothers me to see something be pitched as a basic architectural feature of the clock/time framework when it's so clearly not sufficient.
Has an unsupported return code. Does not change existing behavior (except for fixing the ordering issue). Not a basic architectural feature. Just a common one we want to handle correctly.
Quoted 10 lines
Quoted 6 lines
We just have a difference in where we're trying to re-add the code. I'm moving the current code into the generic tod path, and you're moving it into the RTC driver. Both efforts are consolidating code, so even with the minor duplication we have less code then before. So I'm sure as we whittle away at this we can find a way to meet in the middle. I think this patch moves us forward in that direction.
So long as this GTOD patch is updated to reflect the reality that not all _boards_ can (or will) support that style of RTC ... cleanup is good, but those code comments suggest deeper assumptions.
Sure. I apologize for the confusion.
Quoted 7 lines
Quoted 5 lines
I'm trying to just move us to:
1)timeofday_resume: reset clocksource, NTP, [set xtime]
2) RTC resume: [set xtime]Again, where the [set xtime]s depend on platform config
Getting rid of one is useful, yes. But why not get rid of two?
Sure, but that requires RTC drivers to be written for every arch and platform first.
Quoted 10 lines
Quoted 6 lines
Then as the RTC drivers gain coverage, maybe we can start cutting timeofday resume down and have something like:
1)timeofday_resume: reset clocksource, NTP
2) RTC resume: set xtimeDoes this sound like a way forward?
I'm not sure I see how the last two groups are very different; are you saying it would get rid of these message-unfriendly
read_persistent_clock()calls?
I'm open to that. Once the RTC drivers get there. :)
I apologize if this reply seems a bit terse (and more worrying, combative). That's not the intent. I really do appreciate your review and feedback. Your point about message based RTCs is solid and I will keep it in mind as the development continues.
thanks again,
-john