clocksource cleanup / improvement
Subsystems (4)
cover clocksource cleanup / improvement
Greetings,
I had a go at the clocksource code and tried to make it better. The first 4 patches contain some code cleanup and improve the separation between timekeeping.c and clocksource.[ch].
Patch #5 is the one with the real change: the switch to a new clock source is now done with stop_machine instead of checking with every wall time update. A clocksource change is an extremely rare thing, to use stop_machine seems reasonable to me.
What I'd like to do is to make the content of struct clocksource private to clocksource.c and clocksource.h. The biggest road-block to get there is update_wall_time. I stared at it for hours and then gave up. A bit of unfinished business ..
The clocksource patches have been created against todays upstream tree plus the patches from the tip tree, if anyone wants to try them you need to pull from:
git://git.kernel.org/pub/scm/linux/kernel/git/x86/linux-2.6-tip
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
RFC PATCH 1/5 move clock source related code to clocksource.c
From: Martin Schwidefsky <[email protected]>
Move clock source related code from timekeeping.c to clocksource.c where they belong. The selected clocks source "clock" is now defined in clocksource.c and clocksource_init is added to set up the initial clock.
--- include/linux/clocksource.h | 44 ++------- kernel/time/clocksource.c | 199 ++++++++++++++++++++++++++++++++++++++++++++ kernel/time/timekeeping.c | 161 ----------------------------------- 3 files changed, 211 insertions(+), 193 deletions(-)
include/linux/clocksource.h +11 -33
@@ -14,6 +14,7 @@#include <linux/list.h> #include <linux/cache.h> #include <linux/timer.h> +#include <linux/init.h> #include <asm/div64.h> #include <asm/io.h>
@@ -329,46 +330,23 @@ static inline s64 cyc2ns(struct clocksoureturn ret; } -/** -* clocksource_calculate_interval - Calculates a clocksource interval struct -* -* @c: Pointer to clocksource. -* @length_nsec: Desired interval length in nanoseconds. -* -* Calculates a fixed cycle/nsec interval for a given clocksource/adjustment -* pair and interval request. -* -* Unless you're the timekeeping code, you should not be using this! -*/ -static inline void clocksource_calculate_interval (struct clocksource *c, -unsigned long length_nsec) -{ -u64 tmp; - -/* Do the ns -> cycle conversion first, using original mult */ - tmp = length_nsec; - tmp <<= c->shift; - tmp += c->mult_orig/2 ; -do_div (tmp, c->mult_orig); - - c->cycle_interval = (cycle_t )tmp; -if (c->cycle_interval ==0 ) - c->cycle_interval =1 ; - -/* Go back from cycles -> shifted ns, this time use ntp adjused mult */ - c->xtime_interval = (u64 )c->cycle_interval * c->mult; - c->raw_interval = ((u64 )c->cycle_interval * c->mult_orig) >> c->shift; -} - -/* used to install a new clocksource */ +extern void __init clocksource_init (void );extern int clocksource_register (struct clocksource *);extern void clocksource_unregister (struct clocksource *);extern void clocksource_touch_watchdog (void ); -extern struct clocksource *clocksource_get_next (void );extern void clocksource_change_rating (struct clocksource *cs,int rating); +extern void clocksource_adjust (s64 offset);extern void clocksource_resume (void ); +#ifdef CONFIG_GENERIC_TIME +extern void clocksource_forward_now (void ); +extern void change_clocksource (void ); +#else +static inline void clocksource_forward_now (void ) { } +static inline void change_clocksource (void ) { } +#endif +#ifdef CONFIG_GENERIC_TIME_VSYSCALL extern void update_vsyscall (struct timespec *ts,struct clocksource *c);extern void update_vsyscall_tz (void );
kernel/time/clocksource.c +199 -0
217 lines not on this page · raw message
kernel/time/timekeeping.c +1 -160
189 lines not on this page · raw message
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 7 lines · Martin Schwidefsky
plain text document attachment (clocksource-move.diff)
From: Martin Schwidefsky <[email protected]>Move clock source related code from
timekeeping.ctoclocksource.cwhere they belong. The selected clocks source "clock" is now defined inclocksource.candclocksource_initis added to set up the initial clock.
The problem is most (if not all) the code your moving is actually time keeping code .. The reason it seems like clocksource code is cause John wasn't very choosy about which structure he added variables too .. So really this clean up needs to be in reverse, remove all the timekeeping code from the clocksource code.
Daniel
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 9 lines · On Tue, 21 Jul 2009 12:50:51 -0700 Daniel Walker <[email protected]> wrote:
Quoted 7 lines · Martin Schwidefsky
plain text document attachment (clocksource-move.diff)
From: Martin Schwidefsky <[email protected]>Move clock source related code from
timekeeping.ctoclocksource.cwhere they belong. The selected clocks source "clock" is now defined inclocksource.candclocksource_initis added to set up the initial clock.
Quoted 5 lines
The problem is most (if not all) the code your moving is actually time keeping code .. The reason it seems like clocksource code is cause John wasn't very choosy about which structure he added variables too .. So really this clean up needs to be in reverse, remove all the timekeeping code from the clocksource code.
You are refering to the following variables, aren't you?
cycle_t cycle_interval;
u64 xtime_interval;
u32 raw_interval;
u64 xtime_nsec;
s64 error;
struct timespec raw_time;There is a dependency between these values and the clocksource. But they are only needed for the current active clock, so we could move them out of the clocksource data structure and make them static variables in timekeeping.c. I guess it would improve the code separation for update_wall_time as well. Hmm, worth a try.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
In reply to Daniel WalkerQuoted 9 lines · Daniel Walker
Quoted 7 lines · Martin Schwidefsky
plain text document attachment (clocksource-move.diff)
From: Martin Schwidefsky <[email protected]>Move clock source related code from
timekeeping.ctoclocksource.cwhere they belong. The selected clocks source "clock" is now defined inclocksource.candclocksource_initis added to set up the initial clock.
Quoted 5 lines
The problem is most (if not all) the code your moving is actually time keeping code .. The reason it seems like clocksource code is cause John wasn't very choosy about which structure he added variables too .. So really this clean up needs to be in reverse, remove all the timekeeping code from the clocksource code.
Not so much that I wasn't very choosy, but that I had to pick my battles there. At the time, Roman claimed keeping the timekeeping values in the clocksource (instead of global to timekeeping.c) actually produced better code.
I do agree with Daniel's main point, that the patch mixes the layers I tried to establish in the design.
Clocksource: Abstracts out a hardware counter.
NTP: Provides the reference time.
Timekeeping: Manages accumulating the clocksource, and combining input from ntp's reference time to steer the hardware frequency.
Unfortunately, many timekeeping values got stuffed into the struct clocksource. I've had plans to try to clean this up and utilize Patrick Ohly's simpler clockcounter struct as a basis for a clocksource, nesting the structures somewhat to look something like:
/* minimal structure only giving hardware info and access methods */
struct cyclecounter {
char *name;
cycle_t (*read)(const struct cyclecounter *cc);
cycle_t (*vread)(const struct cyclecounter *cc);
cycle_t mask;
u32 mult;
u32 shift;
};/* more complicated structure holding timekeeping values */
struct timesource {
struct cyclecounter counter;
u32 corrected_mult;
cycle_t cycle_interval;
u64 xtime_interval;
u32 raw_interval;
cycle_t cycle_last;
u64 xtime_nsec;
s64 error; /* probably should be ntp_error */
...
}However such a change would be quite a bit of churn to much of the timekeeping code, and to only marginal benefit. So I've put it off.
Martin, I've not been able to review your changes in extreme detail, but I'm curious what the motivation for the drastic code rearrangement was?
I see you pushing a fair amount of code down a level, for instance, except for the locking, getmonotonicraw() basically gets pushed down to clocksource_read_raw(). The ktime_get/ktime_get_ts/getnstimeofday do reduce some duplicate code, but that could still be minimized without pushing stuff down to the clocksource level.
Is there any streamlining performance benefit from the way you moved things around?
thanks
-john
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 16 lines · On Tue, 21 Jul 2009 15:00:07 -0700 john stultz <[email protected]> wrote:
Quoted 9 lines · Daniel Walker
Quoted 7 lines · Martin Schwidefsky
plain text document attachment (clocksource-move.diff)
From: Martin Schwidefsky <[email protected]>Move clock source related code from
timekeeping.ctoclocksource.cwhere they belong. The selected clocks source "clock" is now defined inclocksource.candclocksource_initis added to set up the initial clock.Quoted 5 lines
The problem is most (if not all) the code your moving is actually time keeping code .. The reason it seems like clocksource code is cause John wasn't very choosy about which structure he added variables too .. So really this clean up needs to be in reverse, remove all the timekeeping code from the clocksource code.
Quoted 4 lines
Not so much that I wasn't very choosy, but that I had to pick my battles there. At the time, Roman claimed keeping the timekeeping values in the clocksource (instead of global to
timekeeping.c) actually produced better code.
There are only a couple of functions where the performance is important, these are the ktime_get functions (and variants) and probably the update_wall_time function.
Quoted 7 lines
I do agree with Daniel's main point, that the patch mixes the layers I tried to establish in the design.
Clocksource: Abstracts out a hardware counter.
NTP: Provides the reference time.
Timekeeping: Manages accumulating the clocksource, and combining input from ntp's reference time to steer the hardware frequency.
Imho what makes the code hard to understand is that the internals of the clocksource have leaked into the timekeeping code. I'm getting at the cycle, mult and shift values here. The code would be much easier to understand if the clocksource would just return nanoseconds. The bad thing here is that we would loose some bits of precision.
Quoted 29 lines
Unfortunately, many timekeeping values got stuffed into the struct clocksource. I've had plans to try to clean this up and utilize Patrick Ohly's simpler clockcounter struct as a basis for a clocksource, nesting the structures somewhat to look something like:
/* minimal structure only giving hardware info and access methods */
struct cyclecounter { char *name; cycle_t (*read)(const struct cyclecounter *cc); cycle_t (*vread)(const struct cyclecounter *cc); cycle_t mask; u32 mult; u32 shift; };/* more complicated structure holding timekeeping values */
struct timesource { struct cyclecounter counter; u32 corrected_mult; cycle_t cycle_interval; u64 xtime_interval; u32 raw_interval; cycle_t cycle_last; u64 xtime_nsec; s64 error; /* probably should be ntp_error */ ... }
Quoted 2 lines
However such a change would be quite a bit of churn to much of the timekeeping code, and to only marginal benefit. So I've put it off.
That would be an improvement, but there are still these pesky cycles in the timesource.
Quoted 2 lines
Martin, I've not been able to review your changes in extreme detail, but I'm curious what the motivation for the drastic code rearrangement was?
It started of with a minor performance optimization, I wanted to get rid of the change_clocksource call every tick. When I looked at the code to understand it I started to move things around.
Quoted 5 lines
I see you pushing a fair amount of code down a level, for instance, except for the locking,
getmonotonicraw()basically gets pushed down toclocksource_read_raw(). The ktime_get/ktime_get_ts/getnstimeofday do reduce some duplicate code, but that could still be minimized without pushing stuff down to the clocksource level.
The background here is that I want to isolate the use ofthe cycles, mult and shift values to clocksource.[ch]
Quoted 2 lines
Is there any streamlining performance benefit from the way you moved things around?
The biggest gain is the ktime_get optimization that is already in the tip tree. The call to change_clocksource showed up in the instructions trace for a cpu wake up and I started to wonder if this is necessary.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 36 lines · Martin Schwidefsky
Quoted 16 lines · On Tue, 21 Jul 2009 15:00:07 -0700 john stultz <[email protected]> wrote:
Quoted 9 lines · Daniel Walker
Quoted 7 lines · Martin Schwidefsky
plain text document attachment (clocksource-move.diff)
From: Martin Schwidefsky <[email protected]>Move clock source related code from
timekeeping.ctoclocksource.cwhere they belong. The selected clocks source "clock" is now defined inclocksource.candclocksource_initis added to set up the initial clock.Quoted 5 lines
The problem is most (if not all) the code your moving is actually time keeping code .. The reason it seems like clocksource code is cause John wasn't very choosy about which structure he added variables too .. So really this clean up needs to be in reverse, remove all the timekeeping code from the clocksource code.
Quoted 4 lines
Not so much that I wasn't very choosy, but that I had to pick my battles there. At the time, Roman claimed keeping the timekeeping values in the clocksource (instead of global to
timekeeping.c) actually produced better code.There are only a couple of functions where the performance is important, these are the
ktime_getfunctions (and variants) and probably theupdate_wall_timefunction.Quoted 7 lines
I do agree with Daniel's main point, that the patch mixes the layers I tried to establish in the design.
Clocksource: Abstracts out a hardware counter.
NTP: Provides the reference time.
Timekeeping: Manages accumulating the clocksource, and combining input from ntp's reference time to steer the hardware frequency.
Quoted 5 lines
Imho what makes the code hard to understand is that the internals of the clocksource have leaked into the timekeeping code. I'm getting at the cycle, mult and shift values here. The code would be much easier to understand if the clocksource would just return nanoseconds. The bad thing here is that we would loose some bits of precision.
While I completely agree the code is hard to understand, I really don't think that pushing that down to clocksource.c will improve things.
As much as you'd prefer it not, I feel the timekeeping code has to deal with cycles. The consistent translation and accumulation of clocksource cycles into nanoseconds is what timekeeping.c is all about.
We already have interfaces that return nanoseconds, they're gensttimeofday, ktime_get, ktime_get_ts.
Quoted 29 lines
Quoted 29 lines
Unfortunately, many timekeeping values got stuffed into the struct clocksource. I've had plans to try to clean this up and utilize Patrick Ohly's simpler clockcounter struct as a basis for a clocksource, nesting the structures somewhat to look something like:
/* minimal structure only giving hardware info and access methods */
struct cyclecounter { char *name; cycle_t (*read)(const struct cyclecounter *cc); cycle_t (*vread)(const struct cyclecounter *cc); cycle_t mask; u32 mult; u32 shift; };/* more complicated structure holding timekeeping values */
struct timesource { struct cyclecounter counter; u32 corrected_mult; cycle_t cycle_interval; u64 xtime_interval; u32 raw_interval; cycle_t cycle_last; u64 xtime_nsec; s64 error; /* probably should be ntp_error */ ... }
Quoted 5 lines
Quoted 2 lines
However such a change would be quite a bit of churn to much of the timekeeping code, and to only marginal benefit. So I've put it off.
That would be an improvement, but there are still these pesky cycles in the timesource.
Again, I think there has to be. Since some portion of the current time is unaccumulated, it is inherently cycles based. The timekeeping core has to decide when to accumulate those cycles into nanoseconds and store them into xtime. In order to do that, the timekeeping code has to have an idea of where the cycle_last value is. Further, for improved precision, and ntp steering, we use the *_interval values to accumulate in chunks.
Quoted 7 lines
Quoted 2 lines
Martin, I've not been able to review your changes in extreme detail, but I'm curious what the motivation for the drastic code rearrangement was?
It started of with a minor performance optimization, I wanted to get rid of the
change_clocksourcecall every tick. When I looked at the code to understand it I started to move things around.
Quoted 8 lines
Quoted 5 lines
I see you pushing a fair amount of code down a level, for instance, except for the locking,
getmonotonicraw()basically gets pushed down toclocksource_read_raw(). The ktime_get/ktime_get_ts/getnstimeofday do reduce some duplicate code, but that could still be minimized without pushing stuff down to the clocksource level.The background here is that I want to isolate the use ofthe cycles, mult and shift values to clocksource.[ch]
Again I do completely agree the code needs to be cleaned up. Unfortunately there's still a split between the GENERIC_TIME and non GENERIC_TIME arches that keeps us from making some cleanups right now. I'm trying to get this all unified (see my arch_gettimeoffset patches), but until we get all the arches moved over, there's some unfortunate uglys we can't get rid of.
If I can find some cycles today, I'll try to take a rough swing at some of the cleanup I mentioned earlier. Probably won't build, but will maybe give you an idea of the direction I'm thinking about, and then you can let me know where you feel its still too complex. Maybe then we can meet in the middle?
thanks
-john
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 35 lines · john stultz
Quoted 34 lines · Martin Schwidefsky
Quoted 31 lines · On Tue, 21 Jul 2009 15:00:07 -0700 john stultz <[email protected]> wrote:
Unfortunately, many timekeeping values got stuffed into the struct clocksource. I've had plans to try to clean this up and utilize Patrick Ohly's simpler clockcounter struct as a basis for a clocksource, nesting the structures somewhat to look something like:
/* minimal structure only giving hardware info and access methods */
struct cyclecounter { char *name; cycle_t (*read)(const struct cyclecounter *cc); cycle_t (*vread)(const struct cyclecounter *cc); cycle_t mask; u32 mult; u32 shift; };/* more complicated structure holding timekeeping values */
struct timesource { struct cyclecounter counter; u32 corrected_mult; cycle_t cycle_interval; u64 xtime_interval; u32 raw_interval; cycle_t cycle_last; u64 xtime_nsec; s64 error; /* probably should be ntp_error */ ... }However such a change would be quite a bit of churn to much of the timekeeping code, and to only marginal benefit. So I've put it off.
[snip]
Quoted 5 lines
If I can find some cycles today, I'll try to take a rough swing at some of the cleanup I mentioned earlier. Probably won't build, but will maybe give you an idea of the direction I'm thinking about, and then you can let me know where you feel its still too complex. Maybe then we can meet in the middle?
Hey Martin,
So here's a really quick swipe at breaking apart the clocksource struct
into a clocksource only portion and a timekeeping portion.
Caveats:
1) This doesn't completely build. The core bits do, but there's still a few left-over issues (see following caveats). Its just here to give you an idea of what I'm thinking about. I'd of course break it up into more manageable chunks before submitting it.
2) Structure names aren't too great right now. Not sure timeclock is what I want to use, probably system_time or something. Will find/replace before the next revision is sent out.
3) I still need to unify the clocksource and cyclecounter structures, as they're basically redundant now.
4) I still need to fix the update_vsyscall code (shouldn't be hard, I didn't want to run through arch code yet).
5) The TSC clocksource uses cycles_last to avoid very slight skew issues (that otherwise would not be noticed). Not sure how to fix that if we're pulling cycles_last (which is purely timekeeping state) out of the clocksource. Will have to think of something.
Other cleanups still out there in the distant future:
1) Once all arches are converted to GENERIC_TIME, we can remove the ifdefs, and cleanup a lot of the more complicated xtime struct manipulation. It will cleanup update_wall_time() nicely.
2) I have a logarithmic accumulation patch to update_wall_time that will remove the need for xtime_cache to be managed and updated. Just have to spend some additional time making sure its bugfree.
3) Once all arches are converted to using read_persistent_clock(), then the arch specific time initialization can be dropped. Removing the majority of direct xtime structure accesses.
4) Then once the remaining direct wall_to_monotonic and xtime accessors are moved to timekeeping.c we can make those both static and embed them into the core timekeeping structure.
But let me know if this patch doesn't achieve most of the cleanup you wanted to see.
thanks
-john
DOES NOT BUILD! FOR REVIEW PURPOSES ONLY!
---
include/linux/clocksource.h +4 -99
@@ -154,8 +154,6 @@ extern u64 timecounter_cyc2time(struct timecounter *tc,* @flags: flags describing special properties * @vread: vsyscall based read * @resume: resume function for the clocksource, if necessary -* @cycle_interval: Used internally by timekeeping core, please ignore. -* @xtime_interval: Used internally by timekeeping core, please ignore. */ struct clocksource {/*
@@ -169,7 +167,6 @@ struct clocksource {void (*disable)(struct clocksource *cs);cycle_t mask;u32 mult; -u32 mult_orig;u32 shift;unsigned long flags;cycle_t (*vread)(void);
@@ -181,19 +178,6 @@ struct clocksource {#define CLKSRC_FSYS_MMIO_SET (mmio, addr) do { } while (0)#endif -/* timekeeping specific data, ignore */ -cycle_t cycle_interval; -u64 xtime_interval; -u32 raw_interval; -/* -* Second part is written at each timer interrupt -* Keep it in a different cache line to dirty no -* more than one cache line. -*/ -cycle_t cycle_last____cacheline_aligned_in_smp ; -u64 xtime_nsec; -s64 error; -struct timespec raw_time;#ifdef CONFIG_CLOCKSOURCE_WATCHDOG /* Watchdog related data, used by the framework */
@@ -202,7 +186,6 @@ struct clocksource {#endif }; -extern struct clocksource *clock;/* current clocksource */ /* * Clock source flags bits::
@@ -267,99 +250,21 @@ static inline u32 clocksource_hz2mult(u32 hz, u32 shift_constant)return (u32 )tmp; } -/** -* clocksource_read: - Access the clocksource's current cycle value -* @cs: pointer to clocksource being read -* -* Uses the clocksource to return the current cycle_t value -*/ -static inline cycle_t clocksource_read (struct clocksource *cs) -{ -return cs->read (cs); -}/** -* clocksource_enable: - enable clocksource -* @cs: pointer to clocksource +* clocksource_cyc2ns - converts clocksource cycles to nanoseconds * -* Enables the specified clocksource. The clocksource callback -* function should start up the hardware and setup mult and field -* members of struct clocksource to reflect hardware capabilities. -*/ -static inline int clocksource_enable (struct clocksource *cs) -{ -int ret =0 ; - -if (cs->enable) - ret = cs->enable (cs); - -/* save mult_orig on enable */ - cs->mult_orig = cs->mult; - -return ret; -} - -/** -* clocksource_disable: - disable clocksource -* @cs: pointer to clocksource -* -* Disables the specified clocksource. The clocksource callback -* function should power down the now unused hardware block to -* save power. -*/ -static inline void clocksource_disable (struct clocksource *cs) -{ -if (cs->disable) - cs->disable (cs); -} - -/** -* cyc2ns - converts clocksource cycles to nanoseconds -* @cs: Pointer to clocksource -* @cycles: Cycles -* -* Uses the clocksource and ntp ajdustment to convert cycle_ts to nanoseconds. +* Converts cycles to nanoseconds, using the given mult and shift. * * XXX - This could use some mult_lxl_ll() asm optimization */ -static inline s64 cyc2ns (struct clocksource *cs,cycle_t cycles) +static inline s64 clocksource_cyc2ns (cycle_t cycles,u32 mult,u32 shift) {u64 ret = (u64 )cycles; - ret = (ret * cs->mult) >> cs->shift; + ret = (ret * mult) >> shift;return ret; } -/** -* clocksource_calculate_interval - Calculates a clocksource interval struct -* -* @c: Pointer to clocksource. -* @length_nsec: Desired interval length in nanoseconds. -* -* Calculates a fixed cycle/nsec interval for a given clocksource/adjustment -* pair and interval request. -* -* Unless you're the timekeeping code, you should not be using this! -*/ -static inline void clocksource_calculate_interval (struct clocksource *c, -unsigned long length_nsec) -{ -u64 tmp; - -/* Do the ns -> cycle conversion first, using original mult */ - tmp = length_nsec; - tmp <<= c->shift; - tmp += c->mult_orig/2 ; -do_div (tmp, c->mult_orig); - - c->cycle_interval = (cycle_t )tmp; -if (c->cycle_interval ==0 ) - c->cycle_interval =1 ; - -/* Go back from cycles -> shifted ns, this time use ntp adjused mult */ - c->xtime_interval = (u64 )c->cycle_interval * c->mult; - c->raw_interval = ((u64 )c->cycle_interval * c->mult_orig) >> c->shift; -} -/* used to install a new clocksource */ extern int clocksource_register (struct clocksource *);
include/linux/time.h +2 -0
@@ -147,6 +147,8 @@ extern struct timespec timespec_trunc(struct timespec t, unsigned gran);extern int timekeeping_valid_for_hres (void );extern void update_wall_time (void );extern void update_xtime_cache (u64 nsec); +extern void timekeeping_leap_insert (int leapsecond); +struct tms ;extern void do_sys_times (struct tms *);
kernel/time/clocksource.c +3 -2
@@ -182,7 +182,8 @@ static void clocksource_watchdog(unsigned long data) resumed =test_and_clear_bit (0 , &watchdog_resumed); wdnow = watchdog->read (watchdog); - wd_nsec =cyc2ns (watchdog, (wdnow - watchdog_last) & watchdog->mask); + wd_nsec =clocksource_cyc2ns ((wdnow - watchdog_last) & watchdog->mask, + watchdog->mult, watchdog->shift); watchdog_last = wdnow;list_for_each_entry_safe (cs, tmp, &watchdog_list, wd_list) {
@@ -209,7 +210,7 @@ static void clocksource_watchdog(unsigned long data) cs->flags |=CLOCK_SOURCE_WATCHDOG ; cs->wd_last = csnow; }else { - cs_nsec =cyc2ns (cs, (csnow - cs->wd_last) & cs->mask); + cs_nsec =clocksource_cyc2ns ((csnow - cs->wd_last) & cs->mask, cs->mult, cs->shift); cs->wd_last = csnow;/* Check the delta. Might remove from the list ! */ clocksource_ratewd (cs, cs_nsec - wd_nsec);
kernel/time/jiffies.c +0 -1
@@ -61,7 +61,6 @@ struct clocksource clocksource_jiffies = { .read = jiffies_read, .mask = 0xffffffff, /*32bits*/ .mult = NSEC_PER_JIFFY << JIFFIES_SHIFT, /* details above */ - .mult_orig = NSEC_PER_JIFFY << JIFFIES_SHIFT, .shift = JIFFIES_SHIFT, };
kernel/time/ntp.c +5 -9
@@ -142,11 +142,11 @@ static void ntp_update_offset(long offset)* Select how the frequency is to be controlled * and in which mode (PLL or FLL). */ - secs = xtime.tv_sec - time_reftime; + secs =get_seconds () - time_reftime;if (unlikely (time_status &STA_FREQHOLD )) secs =0 ; - time_reftime = xtime.tv_sec; + time_reftime =get_seconds (); offset64 = offset; freq_adj = (offset64 * secs) <<
@@ -194,8 +194,7 @@ static enum hrtimer_restart ntp_leap_second(struct hrtimer *timer)case TIME_OK :break ;case TIME_INS : - xtime.tv_sec--; - wall_to_monotonic.tv_sec++; +timekeeping_leap_insert (-1 ); time_state =TIME_OOP ;printk (KERN_NOTICE "Clock: inserting leap second 23:59:60 UTC\n" );
@@ -203,9 +202,8 @@ static enum hrtimer_restart ntp_leap_second(struct hrtimer *timer) res =HRTIMER_RESTART ;break ;case TIME_DEL : - xtime.tv_sec++; +timekeeping_leap_insert (1 ); time_tai--; - wall_to_monotonic.tv_sec--; time_state =TIME_WAIT ;printk (KERN_NOTICE "Clock: deleting leap second 23:59:59 UTC\n" );
@@ -219,8 +217,6 @@ static enum hrtimer_restart ntp_leap_second(struct hrtimer *timer) time_state =TIME_OK ;break ; } -update_vsyscall (&xtime, clock); -write_sequnlock (&xtime_lock);return res;
@@ -371,7 +367,7 @@ static inline void process_adj_status(struct timex *txc, struct timespec *ts)* reference time to current time. */ if (!(time_status &STA_PLL ) && (txc->status &STA_PLL )) - time_reftime = xtime.tv_sec; + time_reftime =get_seconds ();/* only set allowed bits */ time_status &=STA_RONLY ;
kernel/time/timekeeping.c +274 -176
646 lines not on this page · raw message
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 39 lines · On Wed, 22 Jul 2009 17:28:20 -0700 john stultz <[email protected]> wrote:
Hey Martin,
So here's a really quick swipe at breaking apart the clocksource struct
into a clocksource only portion and a timekeeping portion.Caveats:
1) This doesn't completely build. The core bits do, but there's still a few left-over issues (see following caveats). Its just here to give you an idea of what I'm thinking about. I'd of course break it up into more manageable chunks before submitting it.2) Structure names aren't too great right now. Not sure timeclock is what I want to use, probably
system_timeor something. Will find/replace before the next revision is sent out.3) I still need to unify the clocksource and cyclecounter structures, as they're basically redundant now.
4) I still need to fix the
update_vsyscallcode (shouldn't be hard, I didn't want to run through arch code yet).5) The TSC clocksource uses
cycles_lastto avoid very slight skew issues (that otherwise would not be noticed). Not sure how to fix that if we're pullingcycles_last(which is purely timekeeping state) out of the clocksource. Will have to think of something.Other cleanups still out there in the distant future:
1) Once all arches are converted toGENERIC_TIME, we can remove the ifdefs, and cleanup a lot of the more complicated xtime struct manipulation. It will cleanupupdate_wall_time()nicely.2) I have a logarithmic accumulation patch to
update_wall_timethat will remove the need forxtime_cacheto be managed and updated. Just have to spend some additional time making sure its bugfree.3) Once all arches are converted to using
read_persistent_clock(), then the arch specific time initialization can be dropped. Removing the majority of direct xtime structure accesses.
Quoted 7 lines
4) Then once the remaining direct
wall_to_monotonicand xtime accessors are moved totimekeeping.cwe can make those both static and embed them into the core timekeeping structure.But let me know if this patch doesn't achieve most of the cleanup you wanted to see.
Cool, I'll have a look. What I can see right away is that a lot of the changes I tried yesterday are contained in your patch as well. But your patch is more radical, a lot more is (re-)moved from the struct clocksource. That mult_orig goes aways is good :-)
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
In reply to john stultzHi John,
Quoted 10 lines · On Wed, 22 Jul 2009 17:28:20 -0700 john stultz <[email protected]> wrote:
On Wed, 2009-07-22 at 10:45 -0700, john stultz wrote:
Hey Martin,
So here's a really quick swipe at breaking apart the clocksource struct
into a clocksource only portion and a timekeeping portion.Caveats:
1) This doesn't completely build. The core bits do, but there's still a few left-over issues (see following caveats). Its just here to give you an idea of what I'm thinking about. I'd of course break it up into more manageable chunks before submitting it.
Mine does build but for s390 only - and even there only with a hack ;-)
Quoted 3 lines
2) Structure names aren't too great right now. Not sure timeclock is what I want to use, probably
system_timeor something. Will find/replace before the next revision is sent out.
I've picked the name struct timekeeper.
Quoted 2 lines
3) I still need to unify the clocksource and cyclecounter structures, as they're basically redundant now.
I'll leave that up to you.
Quoted 2 lines
4) I still need to fix the
update_vsyscallcode (shouldn't be hard, I didn't want to run through arch code yet).
Done in my version of the patch below.
Quoted 4 lines
5) The TSC clocksource uses
cycles_lastto avoid very slight skew issues (that otherwise would not be noticed). Not sure how to fix that if we're pullingcycles_last(which is purely timekeeping state) out of the clocksource. Will have to think of something.
That is an ugly one. A similar thing exists in the s390 backend where I want to reset the timekeeping to precise values after the clocksource switch from jiffies. The proper solution probably is to allow architectures to override the default clocksource. The jiffies clocksource doesn't make any sense on s390.
Quoted 4 lines
Other cleanups still out there in the distant future:
1) Once all arches are converted toGENERIC_TIME, we can remove the ifdefs, and cleanup a lot of the more complicated xtime struct manipulation. It will cleanupupdate_wall_time()nicely.
Well yes, but that will take a few more arch patches. Until we get there we have to live with the current state of things.
Quoted 3 lines
2) I have a logarithmic accumulation patch to
update_wall_timethat will remove the need forxtime_cacheto be managed and updated. Just have to spend some additional time making sure its bugfree.
Interesting. Getting rid of xtime_cache would be a nice cleanup as well.
Quoted 3 lines
3) Once all arches are converted to using
read_persistent_clock(), then the arch specific time initialization can be dropped. Removing the majority of direct xtime structure accesses.
Only if the read_persistent_clock allows for a better resolution than seconds. With my cputime accounting hat on: seconds are not good enough..
Quoted 3 lines
4) Then once the remaining direct
wall_to_monotonicand xtime accessors are moved totimekeeping.cwe can make those both static and embed them into the core timekeeping structure.
Both should not be accessed at a rate that makes it necessary to read from the values directly. An accessor should be fine I think.
Quoted 2 lines
But let me know if this patch doesn't achieve most of the cleanup you wanted to see.
We are getting there. I wonder if it is really necessary to pull xtime_cache, raw_time, total_sleep_time and timekeeping_suspended into the struct timeclock. I would prefer the semantics that the struct timekeeper / timeclock contains the internal values of the timekeeping code for the currently selected clock source. xtime is not clock specific.
For reference here is the current stack of patches I have on my disk. The stop_machine conversion to install a new clocksource is currently missing.
PRELIMINARY PATCHES, USE AT YOUR OWN RISK.
-------------------------------------------------------------------
Subject: [PATCH] introduce timekeeping_leap_insert
From: john stultz <[email protected]>
--- include/linux/time.h | 1 + kernel/time/ntp.c | 7 ++----- kernel/time/timekeeping.c | 7 +++++++ 3 files changed, 10 insertions(+), 5 deletions(-)
include/linux/time.h +1 -0
@@ -147,6 +147,7 @@ extern struct timespec timespec_trunc(stextern int timekeeping_valid_for_hres (void );extern void update_wall_time (void );extern void update_xtime_cache (u64 nsec); +extern void timekeeping_leap_insert (int leapsecond);struct tms ;extern void do_sys_times (struct tms *);
kernel/time/ntp.c +2 -5
@@ -194,8 +194,7 @@ static enum hrtimer_restart ntp_leap_seccase TIME_OK :break ;case TIME_INS : - xtime.tv_sec--; - wall_to_monotonic.tv_sec++; +timekeeping_leap_insert (-1 ); time_state =TIME_OOP ;printk (KERN_NOTICE "Clock: inserting leap second 23:59:60 UTC\n" );
@@ -203,9 +202,8 @@ static enum hrtimer_restart ntp_leap_sec res =HRTIMER_RESTART ;break ;case TIME_DEL : - xtime.tv_sec++; +timekeeping_leap_insert (1 ); time_tai--; - wall_to_monotonic.tv_sec--; time_state =TIME_WAIT ;printk (KERN_NOTICE "Clock: deleting leap second 23:59:59 UTC\n" );
@@ -219,7 +217,6 @@ static enum hrtimer_restart ntp_leap_sec time_state =TIME_OK ;break ; } -update_vsyscall (&xtime, clock);write_sequnlock (&xtime_lock);
kernel/time/timekeeping.c +7 -0
@@ -58,6 +58,13 @@ void update_xtime_cache(u64 nsec)struct clocksource *clock; +/* must hold xtime_lock */ +void timekeeping_leap_insert (int leapsecond) +{ + xtime.tv_sec += leapsecond; + wall_to_monotonic.tv_sec -= leapsecond; +update_vsyscall (&xtime, clock); +}#ifdef CONFIG_GENERIC_TIME /**
-------------------------------------------------------------------
Subject: [PATCH] remove clocksource inline functions
From: Martin Schwidefsky <[email protected]>
Remove clocksource_read, clocksource_enable and clocksource_disable inline functions. No functional change.
--- include/linux/clocksource.h | 46 -------------------------------------------- kernel/time/timekeeping.c | 32 +++++++++++++++++------------- 2 files changed, 18 insertions(+), 60 deletions(-)
kernel/time/timekeeping.c +18 -14
@@ -79,7 +79,7 @@ static void clocksource_forward_now(voidcycle_t cycle_now, cycle_delta;s64 nsec; - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock); cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; clock->cycle_last = cycle_now;
@@ -114,7 +114,7 @@ void getnstimeofday(struct timespec *ts) *ts = xtime;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -146,7 +146,7 @@ ktime_t ktime_get(void) nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -186,7 +186,7 @@ void ktime_get_ts(struct timespec *ts) tomono = wall_to_monotonic;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -274,16 +274,18 @@ static void change_clocksource(void)clocksource_forward_now (); -if (clocksource_enable (new)) +if (new->enable && ! new->enable (new))return ; +/* save mult_orig on enable */ + new->mult_orig = new->mult; new->raw_time = clock->raw_time; old = clock; clock = new; -clocksource_disable (old); +if (old->disable) + old->disable (old); - clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; clock->xtime_nsec =0 ;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH );
@@ -373,7 +375,7 @@ void getrawmonotonic(struct timespec *ts seq =read_seqbegin (&xtime_lock);/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -435,9 +437,12 @@ void __init timekeeping_init(void)ntp_init (); clock =clocksource_get_next (); -clocksource_enable (clock); +if (clock->enable) + clock->enable (clock); +/* save mult_orig on enable */ + clock->mult_orig = clock->mult;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH ); - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); xtime.tv_sec = sec; xtime.tv_nsec =0 ;
@@ -477,8 +482,7 @@ static int timekeeping_resume(struct sys }update_xtime_cache (0 );/* re-base the last cycle value */ - clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; timekeeping_suspended =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);
@@ -630,7 +634,7 @@ void update_wall_time(void)return ;#ifdef CONFIG_GENERIC_TIME - offset = (clocksource_read (clock) - clock->cycle_last) & clock->mask; + offset = (clock->read (clock) - clock->cycle_last) & clock->mask;#else offset = clock->cycle_interval;#endif
include/linux/clocksource.h +0 -46
52 lines not on this page · raw message
-------------------------------------------------------------------
Subject: [PATCH] cleanup clocksource selection
From: Martin Schwidefsky <[email protected]>
Introduce clocksource_dequeue & clocksource_update and move spinlock calls. clocksource_update does nothing for GENERIC_TIME=n since change_clocksource does nothing as well.
--- kernel/time/clocksource.c | 91 +++++++++++++++++++++++++++------------------- 1 file changed, 55 insertions(+), 36 deletions(-)
kernel/time/clocksource.c +55 -36
156 lines not on this page · raw message
-------------------------------------------------------------------
Subject: [PATCH] introduce struct timekeeper
From: Martin Schwidefsky <[email protected]>
Add struct timekeeper to keep all the internal values timekeeping.c needs in regard to the currently selected clock source. This moves all timekeeping related values out of the struct clocksource.
--- arch/ia64/kernel/time.c | 5 arch/powerpc/kernel/time.c | 5 arch/s390/kernel/time.c | 9 - arch/x86/kernel/vsyscall_64.c | 5 include/linux/clocksource.h | 51 --------- kernel/time/timekeeping.c | 224 +++++++++++++++++++++++++++--------------- 6 files changed, 161 insertions(+), 138 deletions(-)
include/linux/clocksource.h +2 -49
77 lines not on this page · raw message
kernel/time/timekeeping.c +145 -79
490 lines not on this page · raw message
arch/ia64/kernel/time.c +3 -2
@@ -473,7 +473,8 @@ void update_vsyscall_tz(void) { } -void update_vsyscall (struct timespec *wall,struct clocksource *c) +void update_vsyscall(struct timespec *wall, struct clocksource *c,# +cycle_t cycle_last) {unsigned long flags;
@@ -484,7 +485,7 @@ void update_vsyscall(struct timespec *wa fsyscall_gtod_data.clk_mult = c->mult; fsyscall_gtod_data.clk_shift = c->shift; fsyscall_gtod_data.clk_fsys_mmio = c->fsys_mmio; - fsyscall_gtod_data.clk_cycle_last = c->cycle_last; + fsyscall_gtod_data.clk_cycle_last = cycle_last;/* copy kernel time structures */ fsyscall_gtod_data.wall_time.tv_sec = wall->tv_sec;
arch/powerpc/kernel/time.c +3 -2
@@ -802,7 +802,8 @@ static cycle_t timebase_read(struct clocreturn (cycle_t )get_tb (); } -void update_vsyscall (struct timespec *wall_time,struct clocksource *clock) +void update_vsyscall (struct timespec *wall_time,struct clocksource *clock, +cycle_t cycle_last) {u64 t2x, stamp_xsec;
@@ -819,7 +820,7 @@ void update_vsyscall(struct timespec *wa stamp_xsec = (u64 ) xtime.tv_nsec *XSEC_PER_SEC ;do_div (stamp_xsec,1000000000 ); stamp_xsec += (u64 ) xtime.tv_sec *XSEC_PER_SEC ; -update_gtod (clock->cycle_last, stamp_xsec, t2x); +update_gtod (cycle_last, stamp_xsec, t2x); }void update_vsyscall_tz (void )
arch/s390/kernel/time.c +5 -4
@@ -206,7 +206,8 @@ static struct clocksource clocksource_to }; -void update_vsyscall (struct timespec *wall_time,struct clocksource *clock) +void update_vsyscall (struct timespec *wall_time,struct clocksource *clock, +cycle_t cycle_last) {if (clock != &clocksource_tod)return ;
@@ -214,7 +215,7 @@ void update_vsyscall(struct timespec *wa/* Make userspace gettimeofday spin until we're done. */ ++vdso_data->tb_update_count;smp_wmb (); - vdso_data->xtime_tod_stamp = clock->cycle_last; + vdso_data->xtime_tod_stamp = cycle_last; vdso_data->xtime_clock_sec = xtime.tv_sec; vdso_data->xtime_clock_nsec = xtime.tv_nsec; vdso_data->wtom_clock_sec = wall_to_monotonic.tv_sec;
@@ -275,8 +276,8 @@ void __init time_init(void)write_seqlock_irqsave (&xtime_lock, flags); now =get_clock ();tod_to_timeval (now -TOD_UNIX_EPOCH , &xtime); - clocksource_tod.cycle_last = now; - clocksource_tod.raw_time = xtime; +// clocksource_tod.cycle_last = now; +// clocksource_tod.raw_time = xtime; tod_to_timeval (sched_clock_base_cc -TOD_UNIX_EPOCH , &ts);set_normalized_timespec (&wall_to_monotonic, -ts.tv_sec, -ts.tv_nsec);write_sequnlock_irqrestore (&xtime_lock, flags);
arch/x86/kernel/vsyscall_64.c +3 -2
@@ -73,14 +73,15 @@ void update_vsyscall_tz(void)write_sequnlock_irqrestore (&vsyscall_gtod_data.lock, flags); } -void update_vsyscall (struct timespec *wall_time,struct clocksource *clock) +void update_vsyscall (struct timespec *wall_time,struct clocksource *clock, +cycle_t cycle_last) {unsigned long flags;write_seqlock_irqsave (&vsyscall_gtod_data.lock, flags);/* copy vsyscall data */ vsyscall_gtod_data.clock.vread = clock->vread; - vsyscall_gtod_data.clock.cycle_last = clock->cycle_last; + vsyscall_gtod_data.clock.cycle_last = cycle_last; vsyscall_gtod_data.clock.mask = clock->mask; vsyscall_gtod_data.clock.mult = clock->mult; vsyscall_gtod_data.clock.shift = clock->shift;
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 7 lines · Martin Schwidefsky
Quoted 3 lines · On Wed, 22 Jul 2009 17:28:20 -0700 john stultz <[email protected]> wrote:
2) Structure names aren't too great right now. Not sure timeclock is what I want to use, probably
system_timeor something. Will find/replace before the next revision is sent out.I've picked the name struct timekeeper.
Nice. I like this name.
Quoted 10 lines
Quoted 4 lines
5) The TSC clocksource uses
cycles_lastto avoid very slight skew issues (that otherwise would not be noticed). Not sure how to fix that if we're pullingcycles_last(which is purely timekeeping state) out of the clocksource. Will have to think of something.That is an ugly one. A similar thing exists in the s390 backend where I want to reset the timekeeping to precise values after the clocksource switch from jiffies. The proper solution probably is to allow architectures to override the default clocksource. The jiffies clocksource doesn't make any sense on s390.
Yea. Still not sure what a good fix is here. We need some way for the TSC to check if its slightly behind another cpu. Alternatively we could create a flag (SLIGHTLY_UNSYNCED or something) where we then do the quick comparison in the timekeeping code instead of the clocksource read implementation? It would really only be useful for the TSC and would clutter the code up somewhat, so I'm not sure I really like it, but maybe that would work.
Quoted 7 lines
Quoted 3 lines
3) Once all arches are converted to using
read_persistent_clock(), then the arch specific time initialization can be dropped. Removing the majority of direct xtime structure accesses.Only if the
read_persistent_clockallows for a better resolution than seconds. With my cputime accounting hat on: seconds are not good enough..
cputime accounting? read_persistent_clock is only used for time initialization. But yea, it could be extended to return a timespec.
Quoted 6 lines
Quoted 3 lines
4) Then once the remaining direct
wall_to_monotonicand xtime accessors are moved totimekeeping.cwe can make those both static and embed them into the core timekeeping structure.Both should not be accessed at a rate that makes it necessary to read from the values directly. An accessor should be fine I think.
Yea its just a matter of cleaning things up.
Quoted 9 lines
Quoted 2 lines
But let me know if this patch doesn't achieve most of the cleanup you wanted to see.
We are getting there. I wonder if it is really necessary to pull
xtime_cache,raw_time,total_sleep_timeandtimekeeping_suspendedinto the struct timeclock. I would prefer the semantics that the struct timekeeper / timeclock contains the internal values of the timekeeping code for the currently selected clock source. xtime is not clock specific.
Its not necessary. Although I do like the idea of pulling all the values protected by the xtime_lock into one structure so its clear what we're protecting. As it is now, jiffies, ntp state and a whole host of other data is covered by it, so it would be a ways off until all that logic could be dethreaded.
The other reason for putting those values into the timekeeping struct is that they are actually fairly tightly connected with the clocksource state. There is some redundancy: xtime_nsec stores the highres remainder of xtime.tv_nsec, xtime_cache is also mostly duplicate, so hopefully we can combine those once xtime is made static.
Quoted 8 lines
For reference here is the current stack of patches I have on my disk. The
stop_machineconversion to install a new clocksource is currently missing.PRELIMINARY PATCHES, USE AT YOUR OWN RISK.
-------------------------------------------------------------------Subject: [PATCH] introduce
timekeeping_leap_insert
Quoted 8 lines
From: john stultz <[email protected]>
--- include/linux/time.h | 1 + kernel/time/ntp.c | 7 ++----- kernel/time/timekeeping.c | 7 +++++++ 3 files changed, 10 insertions(+), 5 deletions(-)
[snip]
I think this one is pretty easy and should be able to be submitted on the sooner side.
Quoted 14 lines
-------------------------------------------------------------------
Subject: [PATCH] remove clocksource inline functions
From: Martin Schwidefsky <[email protected]>
Remove
clocksource_read,clocksource_enableandclocksource_disableinline functions. No functional change.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Cc Daniel Walker <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>
Quoted 4 lines
--- include/linux/clocksource.h | 46 -------------------------------------------- kernel/time/timekeeping.c | 32 +++++++++++++++++------------- 2 files changed, 18 insertions(+), 60 deletions(-)
[snip]
Again, the outstanding mult_orig fix (I need to pester Thomas or Andrew to push that before 2.6.31 comes out) will collide with this, but overall it looks ok.
Quoted 10 lines
-------------------------------------------------------------------
Subject: [PATCH] cleanup clocksource selection
From: Martin Schwidefsky <[email protected]>
Introduce
clocksource_dequeue&clocksource_updateand move spinlock calls.clocksource_updatedoes nothing for GENERIC_TIME=n sincechange_clocksourcedoes nothing as well.
Quoted 5 lines
Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Cc Daniel Walker <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>
[snip]
Haven't had enough time to closely review or test this, but it looks like an agreeable change.
Quoted 24 lines
-------------------------------------------------------------------
Subject: [PATCH] introduce struct timekeeper
From: Martin Schwidefsky <[email protected]>
Add struct timekeeper to keep all the internal values
timekeeping.cneeds in regard to the currently selected clock source. This moves all timekeeping related values out of the struct clocksource.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Cc Daniel Walker <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- arch/ia64/kernel/time.c | 5 arch/powerpc/kernel/time.c | 5 arch/s390/kernel/time.c | 9 - arch/x86/kernel/vsyscall_64.c | 5 include/linux/clocksource.h | 51 --------- kernel/time/timekeeping.c | 224 +++++++++++++++++++++++++++--------------- 6 files changed, 161 insertions(+), 138 deletions(-)
[snip]
Quoted 8 lines
kernel/time/timekeeping.c +8
-0+/* Structure holding internal timekeeping values. */ +struct timekeeper { +struct clocksource *clock; +cycle_t cycle_interval; +u64 xtime_interval; +u32 raw_interval; +u64 xtime_nsec; +s64 ntp_error;
kernel/time/timekeeping.c line 39
Quoted 10 lines
kernel/time/timekeeping.c +10
-0+int xtime_shift; +int ntp_shift; + +/* +* The following is written at each timer interrupt +* Keep it in a different cache line to dirty no +* more than one cache line. +*/ +cycle_t cycle_last____cacheline_aligned_in_smp ; +};
The cacheline aligned bit probably needs to cover almost everything here (not clock, not the shift values), as it will all be updated each tick.
[snip]
Quoted 105 lines
kernel/time/timekeeping.c +25
-25@@ -564,8 +628,8 @@ static __always_inline int clocksource_b* Now calculate the error in (1 << look_ahead) ticks, but first * remove the single look ahead already included in the error. */ - tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift +1 ); - tick_error -= clock->xtime_interval >>1 ; + tick_error = tick_length >> (timekeeper.ntp_shift +1 ); + tick_error -= timekeeper.xtime_interval >>1 ; error = ((error - tick_error) >> look_ahead) + tick_error;/* Finally calculate the adjustment shift value. */ @@ -590,18 +654,18 @@ static __always_inline int clocksource_b* this is optimized for the most common adjustments of -1,0,1, * for other values we can do a bit more work. */ -static void clocksource_adjust (s64 offset) +static void timekeeping_adjust (s64 offset) { -s64 error, interval = clock->cycle_interval; +s64 error, interval = timekeeper.cycle_interval;int adj; - error = clock->error >> (NTP_SCALE_SHIFT - clock->shift -1 ); + error = timekeeper.ntp_error >> (timekeeper.ntp_shift -1 );if (error > interval) { error >>=2 ;if (likely (error <= interval)) adj =1 ;else - adj =clocksource_bigadjust (error, &interval, &offset); + adj =timekeeping_bigadjust (error, &interval, &offset); }else if (error < -interval) { error >>=2 ;if (likely (error >= -interval)) {@@ -609,15 +673,14 @@ static void clocksource_adjust(s64 offse interval = -interval; offset = -offset; }else - adj =clocksource_bigadjust (error, &interval, &offset); + adj =timekeeping_bigadjust (error, &interval, &offset); }else return ; - clock->mult += adj; - clock->xtime_interval += interval; - clock->xtime_nsec -= offset; - clock->error -= (interval - offset) << - (NTP_SCALE_SHIFT - clock->shift); + timekeeper.clock->mult += adj; + timekeeper.xtime_interval += interval; + timekeeper.xtime_nsec -= offset; + timekeeper.ntp_error -= (interval - offset) << timekeeper.ntp_shift; }/** @@ -627,53 +690,56 @@ static void clocksource_adjust(s64 offse*/ void update_wall_time (void ) { +struct clocksource *clock;cycle_t offset;/* Make sure we're fully resumed: */ if (unlikely (timekeeping_suspended))return ; + clock = timekeeper.clock;#ifdef CONFIG_GENERIC_TIME - offset = (clock->read (clock) - clock->cycle_last) & clock->mask; + offset = (clock->read (clock) - timekeeper.cycle_last) & clock->mask;#else - offset = clock->cycle_interval; + offset = timekeeper.cycle_interval;#endif - clock->xtime_nsec = (s64 )xtime.tv_nsec << clock->shift; + timekeeper.xtime_nsec = (s64 )xtime.tv_nsec << timekeeper.xtime_shift;/* normally this loop will run just once, however in the * case of lost or late ticks, it will accumulate correctly. */ -while (offset >= clock->cycle_interval) { +while (offset >= timekeeper.cycle_interval) {/* accumulate one interval */ - offset -= clock->cycle_interval; - clock->cycle_last += clock->cycle_interval; + offset -= timekeeper.cycle_interval; + timekeeper.cycle_last += timekeeper.cycle_interval; - clock->xtime_nsec += clock->xtime_interval; -if (clock->xtime_nsec >= (u64 )NSEC_PER_SEC << clock->shift) { - clock->xtime_nsec -= (u64 )NSEC_PER_SEC << clock->shift; + timekeeper.xtime_nsec += timekeeper.xtime_interval; +if (timekeeper.xtime_nsec >= (u64 )NSEC_PER_SEC << timekeeper.xtime_shift) { + timekeeper.xtime_nsec -= (u64 )NSEC_PER_SEC << timekeeper.xtime_shift; xtime.tv_sec++;second_overflow (); } - clock->raw_time.tv_nsec += clock->raw_interval; -if (clock->raw_time.tv_nsec >=NSEC_PER_SEC ) { - clock->raw_time.tv_nsec -=NSEC_PER_SEC ; - clock->raw_time.tv_sec++; + raw_time.tv_nsec += timekeeper.raw_interval; +if (raw_time.tv_nsec >=NSEC_PER_SEC ) { + raw_time.tv_nsec -=NSEC_PER_SEC ;
kernel/time/timekeeping.c static void clocksource_adjust(s64 offse · line 733
Quoted 10 lines
kernel/time/timekeeping.c +4
-2+ raw_time.tv_sec++; }/* accumulate error between NTP and clock interval */ - clock->error += tick_length; - clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift); + timekeeper.ntp_error += tick_length; + timekeeper.ntp_error -= timekeeper.xtime_interval << + timekeeper.ntp_shift; }
Just to be super careful, I'd probably separate the introduction of the timekeeper code and the ntp_shift changes into separate patches. Just to keep the amount of change to very complex code down to a more easily review-able amount.
So yea, so I think we're pretty much seeing the same vision here! There's still the outstanding cycles_last access by the TSC clocksource, so let me know what you think about my idea for that I mentioned earlier.
Looks good. If I can get some more time soon I'll try to put all of these patches together for some testing and see what else I can add.
thanks again!
-john
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
Quoted 10 lines · On Fri, 24 Jul 2009 17:08:47 -0700 john stultz <[email protected]> wrote:
Quoted 7 lines · Martin Schwidefsky
Quoted 3 lines · On Wed, 22 Jul 2009 17:28:20 -0700 john stultz <[email protected]> wrote:
2) Structure names aren't too great right now. Not sure timeclock is what I want to use, probably
system_timeor something. Will find/replace before the next revision is sent out.I've picked the name struct timekeeper.
Nice. I like this name.
Agreed then.
Quoted 11 lines
Quoted 10 lines
Quoted 4 lines
5) The TSC clocksource uses
cycles_lastto avoid very slight skew issues (that otherwise would not be noticed). Not sure how to fix that if we're pullingcycles_last(which is purely timekeeping state) out of the clocksource. Will have to think of something.That is an ugly one. A similar thing exists in the s390 backend where I want to reset the timekeeping to precise values after the clocksource switch from jiffies. The proper solution probably is to allow architectures to override the default clocksource. The jiffies clocksource doesn't make any sense on s390.
Quoted 7 lines
Yea. Still not sure what a good fix is here. We need some way for the TSC to check if its slightly behind another cpu. Alternatively we could create a flag (
SLIGHTLY_UNSYNCEDor something) where we then do the quick comparison in the timekeeping code instead of the clocksource read implementation? It would really only be useful for the TSC and would clutter the code up somewhat, so I'm not sure I really like it, but maybe that would work.
For my next version of the patch series I will leave the cycle_last in the struct clocksource. Small, manageable steps, if we find a way to move it then we can do that later on as well.
Quoted 10 lines
Quoted 7 lines
Quoted 3 lines
3) Once all arches are converted to using
read_persistent_clock(), then the arch specific time initialization can be dropped. Removing the majority of direct xtime structure accesses.Only if the
read_persistent_clockallows for a better resolution than seconds. With my cputime accounting hat on: seconds are not good enough..cputime accounting?
read_persistent_clockis only used for time initialization. But yea, it could be extended to return a timespec.
Exactly!
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
4) Then once the remaining direct
wall_to_monotonicand xtime accessors are moved totimekeeping.cwe can make those both static and embed them into the core timekeeping structure.Both should not be accessed at a rate that makes it necessary to read from the values directly. An accessor should be fine I think.
Yea its just a matter of cleaning things up.
Nod.
Quoted 10 lines
Quoted 9 lines
Quoted 2 lines
But let me know if this patch doesn't achieve most of the cleanup you wanted to see.
We are getting there. I wonder if it is really necessary to pull
xtime_cache,raw_time,total_sleep_timeandtimekeeping_suspendedinto the struct timeclock. I would prefer the semantics that the struct timekeeper / timeclock contains the internal values of the timekeeping code for the currently selected clock source. xtime is not clock specific.
Quoted 5 lines
Its not necessary. Although I do like the idea of pulling all the values protected by the
xtime_lockinto one structure so its clear what we're protecting. As it is now, jiffies, ntp state and a whole host of other data is covered by it, so it would be a ways off until all that logic could be dethreaded.
Yea, I guess that makes sense. Gives us something to do in the near future. For now I do what is not too complicated to implement.
Quoted 22 lines
The other reason for putting those values into the timekeeping struct is that they are actually fairly tightly connected with the clocksource state. There is some redundancy:
xtime_nsecstores the highres remainder of xtime.tv_nsec,xtime_cacheis also mostly duplicate, so hopefully we can combine those once xtime is made static.Quoted 8 lines
For reference here is the current stack of patches I have on my disk. The
stop_machineconversion to install a new clocksource is currently missing.PRELIMINARY PATCHES, USE AT YOUR OWN RISK.
-------------------------------------------------------------------Subject: [PATCH] introduce
timekeeping_leap_insertQuoted 8 lines
From: john stultz <[email protected]>
--- include/linux/time.h | 1 + kernel/time/ntp.c | 7 ++----- kernel/time/timekeeping.c | 7 +++++++ 3 files changed, 10 insertions(+), 5 deletions(-)
Quoted 4 lines
[snip]
I think this one is pretty easy and should be able to be submitted on the sooner side.
A good question is who will upstream all this, will you take care of it ?
Quoted 19 lines
Quoted 14 lines
-------------------------------------------------------------------
Subject: [PATCH] remove clocksource inline functions
From: Martin Schwidefsky <[email protected]>
Remove
clocksource_read,clocksource_enableandclocksource_disableinline functions. No functional change.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Cc Daniel Walker <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>Quoted 4 lines
--- include/linux/clocksource.h | 46 -------------------------------------------- kernel/time/timekeeping.c | 32 +++++++++++++++++------------- 2 files changed, 18 insertions(+), 60 deletions(-)
Quoted 5 lines
[snip]
Again, the outstanding
mult_origfix (I need to pester Thomas or Andrew to push that before 2.6.31 comes out) will collide with this, but overall it looks ok.
The conflict with mult_orig is a small issue, I'll work around it. In the end mult_orig will be gone anyway.
Quoted 16 lines
Quoted 10 lines
-------------------------------------------------------------------
Subject: [PATCH] cleanup clocksource selection
From: Martin Schwidefsky <[email protected]>
Introduce
clocksource_dequeue&clocksource_updateand move spinlock calls.clocksource_updatedoes nothing for GENERIC_TIME=n sincechange_clocksourcedoes nothing as well.Quoted 5 lines
Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Cc Daniel Walker <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>
Quoted 4 lines
[snip]
Haven't had enough time to closely review or test this, but it looks like an agreeable change.
Don't know yet, the change to clocksource for the stop_machine rework will be bigger than I though. The clocksource watchdog code changes the rating of the TSC clock from interrupt context. I need to get out of the interrupt context if I want to use stop_machine for the clocksource update. Not nice..
Quoted 14 lines
Quoted 14 lines
kernel/time/timekeeping.c +14
-0+/* Structure holding internal timekeeping values. */ +struct timekeeper { +struct clocksource *clock; +cycle_t cycle_interval; +u64 xtime_interval; +u32 raw_interval; +u64 xtime_nsec; +s64 ntp_error; +int xtime_shift; +int ntp_shift; + +/* +* The following is written at each timer interrupt +* Keep it in a different cache line to dirty no
Quoted 8 lines
kernel/time/timekeeping.c line 39
Quoted 4 lines
kernel/time/timekeeping.c +4
-0+* more than one cache line. +*/ +cycle_t cycle_last____cacheline_aligned_in_smp ; +};The cacheline aligned bit probably needs to cover almost everything here (not clock, not the shift values), as it will all be updated each tick.
cycle_last is back in the clocksource structure for now.
Quoted 109 lines
[snip]
Quoted 108 lines
kernel/time/timekeeping.c +26
-25@@ -564,8 +628,8 @@ static __always_inline int clocksource_b* Now calculate the error in (1 << look_ahead) ticks, but first * remove the single look ahead already included in the error. */ - tick_error = tick_length >> (NTP_SCALE_SHIFT - clock->shift +1 ); - tick_error -= clock->xtime_interval >>1 ; + tick_error = tick_length >> (timekeeper.ntp_shift +1 ); + tick_error -= timekeeper.xtime_interval >>1 ; error = ((error - tick_error) >> look_ahead) + tick_error;/* Finally calculate the adjustment shift value. */ @@ -590,18 +654,18 @@ static __always_inline int clocksource_b* this is optimized for the most common adjustments of -1,0,1, * for other values we can do a bit more work. */ -static void clocksource_adjust (s64 offset) +static void timekeeping_adjust (s64 offset) { -s64 error, interval = clock->cycle_interval; +s64 error, interval = timekeeper.cycle_interval;int adj; - error = clock->error >> (NTP_SCALE_SHIFT - clock->shift -1 ); + error = timekeeper.ntp_error >> (timekeeper.ntp_shift -1 );if (error > interval) { error >>=2 ;if (likely (error <= interval)) adj =1 ;else - adj =clocksource_bigadjust (error, &interval, &offset); + adj =timekeeping_bigadjust (error, &interval, &offset); }else if (error < -interval) { error >>=2 ;if (likely (error >= -interval)) {@@ -609,15 +673,14 @@ static void clocksource_adjust(s64 offse interval = -interval; offset = -offset; }else - adj =clocksource_bigadjust (error, &interval, &offset); + adj =timekeeping_bigadjust (error, &interval, &offset); }else return ; - clock->mult += adj; - clock->xtime_interval += interval; - clock->xtime_nsec -= offset; - clock->error -= (interval - offset) << - (NTP_SCALE_SHIFT - clock->shift); + timekeeper.clock->mult += adj; + timekeeper.xtime_interval += interval; + timekeeper.xtime_nsec -= offset; + timekeeper.ntp_error -= (interval - offset) << timekeeper.ntp_shift; }/** @@ -627,53 +690,56 @@ static void clocksource_adjust(s64 offse*/ void update_wall_time (void ) { +struct clocksource *clock;cycle_t offset;/* Make sure we're fully resumed: */ if (unlikely (timekeeping_suspended))return ; + clock = timekeeper.clock;#ifdef CONFIG_GENERIC_TIME - offset = (clock->read (clock) - clock->cycle_last) & clock->mask; + offset = (clock->read (clock) - timekeeper.cycle_last) & clock->mask;#else - offset = clock->cycle_interval; + offset = timekeeper.cycle_interval;#endif - clock->xtime_nsec = (s64 )xtime.tv_nsec << clock->shift; + timekeeper.xtime_nsec = (s64 )xtime.tv_nsec << timekeeper.xtime_shift;/* normally this loop will run just once, however in the * case of lost or late ticks, it will accumulate correctly. */ -while (offset >= clock->cycle_interval) { +while (offset >= timekeeper.cycle_interval) {/* accumulate one interval */ - offset -= clock->cycle_interval; - clock->cycle_last += clock->cycle_interval; + offset -= timekeeper.cycle_interval; + timekeeper.cycle_last += timekeeper.cycle_interval; - clock->xtime_nsec += clock->xtime_interval; -if (clock->xtime_nsec >= (u64 )NSEC_PER_SEC << clock->shift) { - clock->xtime_nsec -= (u64 )NSEC_PER_SEC << clock->shift; + timekeeper.xtime_nsec += timekeeper.xtime_interval; +if (timekeeper.xtime_nsec >= (u64 )NSEC_PER_SEC << timekeeper.xtime_shift) { + timekeeper.xtime_nsec -= (u64 )NSEC_PER_SEC << timekeeper.xtime_shift; xtime.tv_sec++;second_overflow (); } - clock->raw_time.tv_nsec += clock->raw_interval; -if (clock->raw_time.tv_nsec >=NSEC_PER_SEC ) { - clock->raw_time.tv_nsec -=NSEC_PER_SEC ; - clock->raw_time.tv_sec++; + raw_time.tv_nsec += timekeeper.raw_interval; +if (raw_time.tv_nsec >=NSEC_PER_SEC ) { + raw_time.tv_nsec -=NSEC_PER_SEC ; + raw_time.tv_sec++; }
Quoted 13 lines
kernel/time/timekeeping.c static void clocksource_adjust(s64 offse · line 733
Quoted 7 lines
kernel/time/timekeeping.c +3
-2/* accumulate error between NTP and clock interval */ - clock->error += tick_length; - clock->error -= clock->xtime_interval << (NTP_SCALE_SHIFT - clock->shift); + timekeeper.ntp_error += tick_length; + timekeeper.ntp_error -= timekeeper.xtime_interval << + timekeeper.ntp_shift; }Just to be super careful, I'd probably separate the introduction of the timekeeper code and the
ntp_shiftchanges into separate patches. Just to keep the amount of change to very complex code down to a more easily review-able amount.
Done, I now have multiple patches for the cleanup. I will resend soon.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 1/5] move clock source related code to clocksource.c
In reply to john stultzQuoted 20 lines · On Wed, 22 Jul 2009 10:45:33 -0700 john stultz <[email protected]> wrote:
Quoted 10 lines · Martin Schwidefsky
Quoted 7 lines · On Tue, 21 Jul 2009 15:00:07 -0700 john stultz <[email protected]> wrote:
I do agree with Daniel's main point, that the patch mixes the layers I tried to establish in the design.
Clocksource: Abstracts out a hardware counter.
NTP: Provides the reference time.
Timekeeping: Manages accumulating the clocksource, and combining input from ntp's reference time to steer the hardware frequency.Quoted 5 lines
Imho what makes the code hard to understand is that the internals of the clocksource have leaked into the timekeeping code. I'm getting at the cycle, mult and shift values here. The code would be much easier to understand if the clocksource would just return nanoseconds. The bad thing here is that we would loose some bits of precision.
While I completely agree the code is hard to understand, I really don't think that pushing that down to
clocksource.cwill improve things.
Quoted 6 lines
As much as you'd prefer it not, I feel the timekeeping code has to deal with cycles. The consistent translation and accumulation of clocksource cycles into nanoseconds is what
timekeeping.cis all about.We already have interfaces that return nanoseconds, they're gensttimeofday,
ktime_get,ktime_get_ts.
After playing around with the idea move some fields from the struct clocksource to a need private structure in timekeeping.c I now agree. The new structure I have in mind currently looks like this:
/* Structure holding internal timekeeping values. */
struct timekeeper {
cycle_t cycle_interval;
u64 xtime_interval;
u32 raw_interval;
u64 xtime_nsec;
s64 ntp_error;
int xtime_shift;
int ntp_shift;
};The raw_time stays in struct clocksource.
Quoted 35 lines
Quoted 29 lines
Quoted 29 lines
Unfortunately, many timekeeping values got stuffed into the struct clocksource. I've had plans to try to clean this up and utilize Patrick Ohly's simpler clockcounter struct as a basis for a clocksource, nesting the structures somewhat to look something like:
/* minimal structure only giving hardware info and access methods */
struct cyclecounter { char *name; cycle_t (*read)(const struct cyclecounter *cc); cycle_t (*vread)(const struct cyclecounter *cc); cycle_t mask; u32 mult; u32 shift; };/* more complicated structure holding timekeeping values */
struct timesource { struct cyclecounter counter; u32 corrected_mult; cycle_t cycle_interval; u64 xtime_interval; u32 raw_interval; cycle_t cycle_last; u64 xtime_nsec; s64 error; /* probably should be ntp_error */ ... }Quoted 5 lines
Quoted 2 lines
However such a change would be quite a bit of churn to much of the timekeeping code, and to only marginal benefit. So I've put it off.
That would be an improvement, but there are still these pesky cycles in the timesource.
Quoted 7 lines
Again, I think there has to be. Since some portion of the current time is unaccumulated, it is inherently cycles based. The timekeeping core has to decide when to accumulate those cycles into nanoseconds and store them into xtime. In order to do that, the timekeeping code has to have an idea of where the
cycle_lastvalue is. Further, for improved precision, and ntp steering, we use the*_intervalvalues to accumulate in chunks.
Yes, I now agree.
Quoted 24 lines
Quoted 7 lines
Quoted 2 lines
Martin, I've not been able to review your changes in extreme detail, but I'm curious what the motivation for the drastic code rearrangement was?
It started of with a minor performance optimization, I wanted to get rid of the
change_clocksourcecall every tick. When I looked at the code to understand it I started to move things around.Quoted 8 lines
Quoted 5 lines
I see you pushing a fair amount of code down a level, for instance, except for the locking,
getmonotonicraw()basically gets pushed down toclocksource_read_raw(). The ktime_get/ktime_get_ts/getnstimeofday do reduce some duplicate code, but that could still be minimized without pushing stuff down to the clocksource level.The background here is that I want to isolate the use ofthe cycles, mult and shift values to clocksource.[ch]
Again I do completely agree the code needs to be cleaned up. Unfortunately there's still a split between the
GENERIC_TIMEand nonGENERIC_TIMEarches that keeps us from making some cleanups right now. I'm trying to get this all unified (see myarch_gettimeoffsetpatches), but until we get all the arches moved over, there's some unfortunate uglys we can't get rid of.
Quoted 5 lines
If I can find some cycles today, I'll try to take a rough swing at some of the cleanup I mentioned earlier. Probably won't build, but will maybe give you an idea of the direction I'm thinking about, and then you can let me know where you feel its still too complex. Maybe then we can meet in the middle?
I'm already in the middle of doing what you suggested. I'll send an update soonish.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
RFC PATCH 2/5 cleanup clocksource selection
In reply to Martin SchwidefskyFrom: Martin Schwidefsky <[email protected]>
Introduce clocksource_dequeue & clocksource_update and move spinlock calls. clocksource_update does nothing for GENERIC_TIME=n since change_clocksource does nothing as well.
--- kernel/time/clocksource.c | 123 ++++++++++++++++++++++++++-------------------- 1 file changed, 70 insertions(+), 53 deletions(-)
kernel/time/clocksource.c +70 -53
@@ -341,46 +341,25 @@ struct clocksource *clocksource_get_nextreturn curr_clocksource; } -/** -* select_clocksource - Selects the best registered clocksource. -* -* Private function. Must hold clocksource_lock when called. -* -* Select the clocksource with the best rating, or the clocksource, -* which is selected by userspace override. -*/ -static struct clocksource *select_clocksource (void ) -{ -struct clocksource *next; - -if (list_empty (&clocksource_list)) -return NULL ; - -if (clocksource_override) - next = clocksource_override; -else - next = list_entry(clocksource_list.next, struct clocksource, - list); - -if (next == curr_clocksource) -return NULL ; - -return next; -} -/* * Enqueue the clocksource sorted by rating */ static int clocksource_enqueue (struct clocksource *c) {struct list_head *tmp, *entry = &clocksource_list; +unsigned long flags; +int rc; +spin_lock_irqsave (&clocksource_lock, flags); + rc =0 ;list_for_each (tmp, &clocksource_list) {struct clocksource *cs; cs = list_entry(tmp, struct clocksource, list); -if (cs == c) -return -EBUSY ; +if (cs == c) { + rc = -EBUSY ; + goto out; + }/* Keep track of the place, where to insert */ if (cs->rating >= c->rating) entry = tmp;
@@ -390,8 +369,23 @@ static int clocksource_enqueue(struct clif (strlen (c->name) ==strlen (override_name) && !strcmp (c->name, override_name)) clocksource_override = c; +out: +spin_unlock_irqrestore (&clocksource_lock, flags); +return rc; +} + +/* +* Dequeue a clocksource +*/ +static void clocksource_dequeue (struct clocksource *cs) +{ +unsigned long flags; -return 0 ; +spin_lock_irqsave (&clocksource_lock, flags); +list_del (&cs->list); +if (clocksource_override == cs) + clocksource_override =NULL ; +spin_unlock_irqrestore (&clocksource_lock, flags); }/**
@@ -539,6 +533,25 @@ void clocksource_forward_now(void) }/** +* select_clocksource - Selects the best registered clocksource. +* +* Private function. Must hold clocksource_lock when called. +* +* Select the clocksource with the best rating, or the clocksource, +* which is selected by userspace override. +*/ +static struct clocksource *select_clocksource (void ) +{ +if (list_empty (&clocksource_list)) +return NULL ; + +if (clocksource_override) +return clocksource_override; + + return list_entry(clocksource_list.next, struct clocksource, list); +} + +/** * change_clocksource - Swaps clocksources if a new one is available * * Accumulates current time interval and initializes new clocksource
@@ -577,6 +590,23 @@ void change_clocksource(void)clock->name); */ } + +/** +* clocksource_update - Check if a better clocksource is available +*/ +static void clocksource_update (void ) +{ +struct clocksource *new; +unsigned long flags; + +spin_lock_irqsave (&clocksource_lock, flags); + new =select_clocksource (); +if (new) + next_clocksource = new; +spin_unlock_irqrestore (&clocksource_lock, flags); +} +#else /* CONFIG_GENERIC_TIME */ +static inline void clocksource_update (void ) { }#endif /**
@@ -587,16 +617,13 @@ void change_clocksource(void)*/ int clocksource_register (struct clocksource *c) { -unsigned long flags;int ret; -spin_lock_irqsave (&clocksource_lock, flags); ret =clocksource_enqueue (c); -if (!ret) - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); -if (!ret) +if (!ret) { +clocksource_update ();clocksource_check_watchdog (c); + }return ret; }EXPORT_SYMBOL (clocksource_register);
@@ -607,14 +634,10 @@ EXPORT_SYMBOL(clocksource_register);*/ void clocksource_change_rating (struct clocksource *cs,int rating) { -unsigned long flags; - -spin_lock_irqsave (&clocksource_lock, flags); -list_del (&cs->list); +clocksource_dequeue (cs); cs->rating = rating;clocksource_enqueue (cs); - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); +clocksource_update (); }/**
@@ -622,14 +645,8 @@ void clocksource_change_rating(struct cl*/ void clocksource_unregister (struct clocksource *cs) { -unsigned long flags; - -spin_lock_irqsave (&clocksource_lock, flags); -list_del (&cs->list); -if (clocksource_override == cs) - clocksource_override =NULL ; - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); +clocksource_dequeue (cs); +clocksource_update (); }/**
@@ -721,13 +738,13 @@ static ssize_t sysfs_override_clocksourc }/* Reselect, when the override name has changed */ -if (ovr != clocksource_override) { +if (ovr != clocksource_override) clocksource_override = ovr; - next_clocksource =select_clocksource (); - }spin_unlock_irq (&clocksource_lock); +clocksource_update (); +return ret; }
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 2/5] cleanup clocksource selection
Quoted 11 lines · Martin Schwidefsky
plain text document attachment (clocksource-queue.diff)
From: Martin Schwidefsky <[email protected]>Introduce
clocksource_dequeue&clocksource_updateand move spinlock calls.clocksource_updatedoes nothing for GENERIC_TIME=n sincechange_clocksourcedoes nothing as well.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>
Quoted 6 lines
--- kernel/time/clocksource.c | 123 ++++++++++++++++++++++++++-------------------- 1 file changed, 70 insertions(+), 53 deletions(-)From a very quick skim, this patch looks structurally ok to me.
thanks
-john
Quoted 206 lines
Index:
linux-2.6/kernel/time/clocksource.c
===================================================================kernel/time/clocksource.c +70
-53@@ -341,46 +341,25 @@ struct clocksource *clocksource_get_nextreturn curr_clocksource; } -/** -* select_clocksource - Selects the best registered clocksource. -* -* Private function. Must hold clocksource_lock when called. -* -* Select the clocksource with the best rating, or the clocksource, -* which is selected by userspace override. -*/ -static struct clocksource *select_clocksource (void ) -{ -struct clocksource *next; - -if (list_empty (&clocksource_list)) -return NULL ; - -if (clocksource_override) - next = clocksource_override; -else - next = list_entry(clocksource_list.next, struct clocksource, - list); - -if (next == curr_clocksource) -return NULL ; - -return next; -} -/* * Enqueue the clocksource sorted by rating */ static int clocksource_enqueue (struct clocksource *c) {struct list_head *tmp, *entry = &clocksource_list; +unsigned long flags; +int rc; +spin_lock_irqsave (&clocksource_lock, flags); + rc =0 ;list_for_each (tmp, &clocksource_list) {struct clocksource *cs; cs = list_entry(tmp, struct clocksource, list); -if (cs == c) -return -EBUSY ; +if (cs == c) { + rc = -EBUSY ; + goto out; + }/* Keep track of the place, where to insert */ if (cs->rating >= c->rating) entry = tmp;@@ -390,8 +369,23 @@ static int clocksource_enqueue(struct clif (strlen (c->name) ==strlen (override_name) && !strcmp (c->name, override_name)) clocksource_override = c; +out: +spin_unlock_irqrestore (&clocksource_lock, flags); +return rc; +} + +/* +* Dequeue a clocksource +*/ +static void clocksource_dequeue (struct clocksource *cs) +{ +unsigned long flags; -return 0 ; +spin_lock_irqsave (&clocksource_lock, flags); +list_del (&cs->list); +if (clocksource_override == cs) + clocksource_override =NULL ; +spin_unlock_irqrestore (&clocksource_lock, flags); }/** @@ -539,6 +533,25 @@ void clocksource_forward_now(void) }/** +* select_clocksource - Selects the best registered clocksource. +* +* Private function. Must hold clocksource_lock when called. +* +* Select the clocksource with the best rating, or the clocksource, +* which is selected by userspace override. +*/ +static struct clocksource *select_clocksource (void ) +{ +if (list_empty (&clocksource_list)) +return NULL ; + +if (clocksource_override) +return clocksource_override; + + return list_entry(clocksource_list.next, struct clocksource, list); +} + +/** * change_clocksource - Swaps clocksources if a new one is available * * Accumulates current time interval and initializes new clocksource @@ -577,6 +590,23 @@ void change_clocksource(void)clock->name); */ } + +/** +* clocksource_update - Check if a better clocksource is available +*/ +static void clocksource_update (void ) +{ +struct clocksource *new; +unsigned long flags; + +spin_lock_irqsave (&clocksource_lock, flags); + new =select_clocksource (); +if (new) + next_clocksource = new; +spin_unlock_irqrestore (&clocksource_lock, flags); +} +#else /* CONFIG_GENERIC_TIME */ +static inline void clocksource_update (void ) { }#endif /** @@ -587,16 +617,13 @@ void change_clocksource(void)*/ int clocksource_register (struct clocksource *c) { -unsigned long flags;int ret; -spin_lock_irqsave (&clocksource_lock, flags); ret =clocksource_enqueue (c); -if (!ret) - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); -if (!ret) +if (!ret) { +clocksource_update ();clocksource_check_watchdog (c); + }return ret; }EXPORT_SYMBOL (clocksource_register);@@ -607,14 +634,10 @@ EXPORT_SYMBOL(clocksource_register);*/ void clocksource_change_rating (struct clocksource *cs,int rating) { -unsigned long flags; - -spin_lock_irqsave (&clocksource_lock, flags); -list_del (&cs->list); +clocksource_dequeue (cs); cs->rating = rating;clocksource_enqueue (cs); - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); +clocksource_update (); }/** @@ -622,14 +645,8 @@ void clocksource_change_rating(struct cl*/ void clocksource_unregister (struct clocksource *cs) { -unsigned long flags; - -spin_lock_irqsave (&clocksource_lock, flags); -list_del (&cs->list); -if (clocksource_override == cs) - clocksource_override =NULL ; - next_clocksource =select_clocksource (); -spin_unlock_irqrestore (&clocksource_lock, flags); +clocksource_dequeue (cs); +clocksource_update (); }/** @@ -721,13 +738,13 @@ static ssize_t sysfs_override_clocksourc }/* Reselect, when the override name has changed */ -if (ovr != clocksource_override) { +if (ovr != clocksource_override) clocksource_override = ovr; - next_clocksource =select_clocksource (); - }spin_unlock_irq (&clocksource_lock); +clocksource_update (); +return ret; }
RFC PATCH 3/5 remove clocksource inline functions
In reply to Martin SchwidefskyFrom: Martin Schwidefsky <[email protected]>
Remove clocksource_read, clocksource_enable and clocksource_disable inline functions. No functional change.
--- kernel/time/clocksource.c | 18 ++++++++++++------ kernel/time/timekeeping.c | 13 ++++++------- 2 files changed, 18 insertions(+), 13 deletions(-)
kernel/time/clocksource.c +12 -6
@@ -517,7 +517,7 @@ void clocksource_forward_now(void)cycle_t cycle_now, cycle_delta;s64 nsec; - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock); cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; clock->cycle_last = cycle_now;
@@ -567,16 +567,19 @@ void change_clocksource(void)clocksource_forward_now (); -if (clocksource_enable (new)) +if (new->enable && !new->enable (new))return ; +/* save mult_orig after successful enable */ + new->mult_orig = new->mult; + new->raw_time = clock->raw_time; old = clock; clock = new; -clocksource_disable (old); +if (old->disable) + old->disable (old); - clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; clock->xtime_nsec =0 ;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH );
@@ -655,7 +658,10 @@ void clocksource_unregister(struct clockvoid __init clocksource_init (void ) { clock =clocksource_get_next (); -clocksource_enable (clock); +if (clock->enable) + clock->enable (clock); +/* save mult_orig on enable */ + clock->mult_orig = clock->mult;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH ); clock->cycle_last = clock->read (clock); }
kernel/time/timekeeping.c +6 -7
@@ -78,7 +78,7 @@ void getnstimeofday(struct timespec *ts) *ts = xtime;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -110,7 +110,7 @@ ktime_t ktime_get(void) nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -150,7 +150,7 @@ void ktime_get_ts(struct timespec *ts) tomono = wall_to_monotonic;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -296,7 +296,7 @@ void getrawmonotonic(struct timespec *ts seq =read_seqbegin (&xtime_lock);/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
@@ -396,8 +396,7 @@ static int timekeeping_resume(struct sys }update_xtime_cache (0 );/* re-base the last cycle value */ - clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; timekeeping_suspended =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);
@@ -464,7 +463,7 @@ void update_wall_time(void)return ;#ifdef CONFIG_GENERIC_TIME - offset = (clocksource_read (clock) - clock->cycle_last) & clock->mask; + offset = (clock->read (clock) - clock->cycle_last) & clock->mask;#else offset = clock->cycle_interval;#endif
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 3/5] remove clocksource inline functions
Quoted 20 lines · Martin Schwidefsky
plain text document attachment (clocksource-inline.diff)
From: Martin Schwidefsky <[email protected]>Remove
clocksource_read,clocksource_enableandclocksource_disableinline functions. No functional change.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- kernel/time/clocksource.c | 18 ++++++++++++------ kernel/time/timekeeping.c | 13 ++++++------- 2 files changed, 18 insertions(+), 13 deletions(-)Index:
linux-2.6/kernel/time/clocksource.c
===================================================================
kernel/time/clocksource.c void clocksource_forward_now(void) · line 520
Quoted 5 lines
kernel/time/clocksource.c +1
-1@@ -517,7 +517,7 @@ void clocksource_forward_now(void)cycle_t cycle_now, cycle_delta;s64 nsec; - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);
What the benefit of this? You get the same result either way. I think it's actually less error prone (less confusing) using the inlines ..
Daniel
Re: [RFC][patch 3/5] remove clocksource inline functions
In reply to Martin SchwidefskyQuoted 6 lines · Martin Schwidefsky
plain text document attachment (clocksource-inline.diff)
From: Martin Schwidefsky <[email protected]>Remove
clocksource_read,clocksource_enableandclocksource_disableinline functions. No functional change.
Quoted 8 lines
Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- kernel/time/clocksource.c | 18 ++++++++++++------ kernel/time/timekeeping.c | 13 ++++++------- 2 files changed, 18 insertions(+), 13 deletions(-)
There's a patch Magnus has that (should be) queued for 2.6.31 that makes unbreaks the mult_orig manipulations enable/disable does. I suspect this will cause this inline function removing to not be that much of a benefit.
thanks
-john
Quoted 20 lines
Index:
linux-2.6/kernel/time/clocksource.c
===================================================================kernel/time/clocksource.c +2
-2@@ -517,7 +517,7 @@ void clocksource_forward_now(void)cycle_t cycle_now, cycle_delta;s64 nsec; - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock); cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; clock->cycle_last = cycle_now;@@ -567,16 +567,19 @@ void change_clocksource(void)clocksource_forward_now (); -if (clocksource_enable (new)) +if (new->enable && !new->enable (new))return ;
kernel/time/clocksource.c void change_clocksource(void) · line 580
Quoted 10 lines
kernel/time/clocksource.c +5
-1+/* save mult_orig after successful enable */ + new->mult_orig = new->mult; + new->raw_time = clock->raw_time; old = clock; clock = new; -clocksource_disable (old); +if (old->disable) + old->disable (old);
mult_orig needs to be saved to mult at this point.
Quoted 84 lines
kernel/time/clocksource.c +5
-3- clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; clock->xtime_nsec =0 ;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH );@@ -655,7 +658,10 @@ void clocksource_unregister(struct clockvoid __init clocksource_init (void ) { clock =clocksource_get_next (); -clocksource_enable (clock); +if (clock->enable) + clock->enable (clock); +/* save mult_orig on enable */ + clock->mult_orig = clock->mult;clocksource_calculate_interval (clock,NTP_INTERVAL_LENGTH ); clock->cycle_last = clock->read (clock); }Index:
linux-2.6/kernel/time/timekeeping.c
===================================================================kernel/time/timekeeping.c +6
-7@@ -78,7 +78,7 @@ void getnstimeofday(struct timespec *ts) *ts = xtime;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;@@ -110,7 +110,7 @@ ktime_t ktime_get(void) nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;@@ -150,7 +150,7 @@ void ktime_get_ts(struct timespec *ts) tomono = wall_to_monotonic;/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;@@ -296,7 +296,7 @@ void getrawmonotonic(struct timespec *ts seq =read_seqbegin (&xtime_lock);/* read clocksource: */ - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock);/* calculate the delta since the last update_wall_time: */ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;@@ -396,8 +396,7 @@ static int timekeeping_resume(struct sys }update_xtime_cache (0 );/* re-base the last cycle value */ - clock->cycle_last =0 ; - clock->cycle_last =clocksource_read (clock); + clock->cycle_last = clock->read (clock); clock->error =0 ; timekeeping_suspended =0 ;write_sequnlock_irqrestore (&xtime_lock, flags);@@ -464,7 +463,7 @@ void update_wall_time(void)return ;#ifdef CONFIG_GENERIC_TIME - offset = (clocksource_read (clock) - clock->cycle_last) & clock->mask; + offset = (clock->read (clock) - clock->cycle_last) & clock->mask;#else offset = clock->cycle_interval;#endif Signature · 5 lines
-- blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 3/5] remove clocksource inline functions
Quoted 17 lines · On Tue, 21 Jul 2009 15:03:55 -0700 john stultz <[email protected]> wrote:
Quoted 6 lines · Martin Schwidefsky
plain text document attachment (clocksource-inline.diff)
From: Martin Schwidefsky <[email protected]>Remove
clocksource_read,clocksource_enableandclocksource_disableinline functions. No functional change.Quoted 8 lines
Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- kernel/time/clocksource.c | 18 ++++++++++++------ kernel/time/timekeeping.c | 13 ++++++------- 2 files changed, 18 insertions(+), 13 deletions(-)
Quoted 4 lines
There's a patch Magnus has that (should be) queued for 2.6.31 that makes unbreaks the
mult_origmanipulations enable/disable does. I suspect this will cause this inline function removing to not be that much of a benefit.
After moving the enable/disable code over to clocksource.c there is no reason for the inline function anymore, no?
Quoted 21 lines
Quoted 21 lines
Index:
linux-2.6/kernel/time/clocksource.c
===================================================================kernel/time/clocksource.c +2
-2@@ -517,7 +517,7 @@ void clocksource_forward_now(void)cycle_t cycle_now, cycle_delta;s64 nsec; - cycle_now =clocksource_read (clock); + cycle_now = clock->read (clock); cycle_delta = (cycle_now - clock->cycle_last) & clock->mask; clock->cycle_last = cycle_now;@@ -567,16 +567,19 @@ void change_clocksource(void)clocksource_forward_now (); -if (clocksource_enable (new)) +if (new->enable && !new->enable (new))return ;
Quoted 11 lines
kernel/time/clocksource.c void change_clocksource(void) · line 580
Quoted 9 lines
kernel/time/clocksource.c +5
-1+/* save mult_orig after successful enable */ + new->mult_orig = new->mult; + new->raw_time = clock->raw_time; old = clock; clock = new; -clocksource_disable (old); +if (old->disable) + old->disable (old);
mult_origneeds to be saved to mult at this point.
Yes, it would make sense to do that. The old code doesn't so I did not want to add it in a cleanup patch. That should be a patch on its own.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
RFC PATCH 4/5 clocksource_read/clocksource_read_raw inline functions
In reply to Martin SchwidefskyFrom: Martin Schwidefsky <[email protected]>
Add clocksource_read / clocksource_read_raw inline functions and use them for getnstimeofday, ktime_get, ktime_get_ts and getrawmonotonic.
--- include/linux/clocksource.h | 45 +++++++++++++++++++++++++++++++++++++++++++ kernel/time/timekeeping.c | 46 +++----------------------------------------- 2 files changed, 49 insertions(+), 42 deletions(-)
include/linux/clocksource.h +45 -0
@@ -284,6 +284,51 @@ static inline s64 cyc2ns(struct clocksoureturn ret; } +/** +* clocksource_read: - Read nanosecond delta from clocksource. +* @cs: pointer to clocksource being read +* +* Read from the clock source and return the clock value converted +* to nanoseconds. +*/ +static inline s64 clocksource_read (struct clocksource *cs) +{ +cycle_t cycle_now, cycle_delta; + +/* read clocksource: */ + cycle_now = cs->read (cs); + +/* calculate the delta since the last update_wall_time: */ + cycle_delta = (cycle_now - cs->cycle_last) & cs->mask; + +/* convert to nanoseconds: */ +return cyc2ns (cs, cycle_delta); +} + +/** +* clocksource_read_raw: - Read raw nanosecond delta from clocksource. +* @cs: pointer to clocksource being read +* +* Read from the clock source and return the clock value converted +* to nanoseconds. +*/ +static inline s64 clocksource_read_raw (struct clocksource *cs, +struct timespec *ts) +{ +cycle_t cycle_now, cycle_delta; + + *ts = clock->raw_time; + +/* read clocksource: */ + cycle_now = cs->read (cs); + +/* calculate the delta since the last update_wall_time: */ + cycle_delta = (cycle_now - cs->cycle_last) & cs->mask; + +/* convert to nanoseconds: */ +return ((s64 ) cycle_delta * clock->mult_orig) >> clock->shift; +} +/* used to install a new clocksource */ extern void __init clocksource_init (void );extern int clocksource_register (struct clocksource *);
kernel/time/timekeeping.c +4 -42
91 lines not on this page · raw message
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
Re: [RFC][patch 4/5] clocksource_read/clocksource_read_raw inline functions
Quoted 33 lines · Martin Schwidefsky
plain text document attachment (clocksource-read-ns.diff)
From: Martin Schwidefsky <[email protected]>Add
clocksource_read/clocksource_read_rawinline functions and use them for getnstimeofday,ktime_get,ktime_get_tsand getrawmonotonic.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- include/linux/clocksource.h | 45 +++++++++++++++++++++++++++++++++++++++++++ kernel/time/timekeeping.c | 46 +++----------------------------------------- 2 files changed, 49 insertions(+), 42 deletions(-)Index:
linux-2.6/include/linux/clocksource.h
===================================================================include/linux/clocksource.h +10
-0@@ -284,6 +284,51 @@ static inline s64 cyc2ns(struct clocksoureturn ret; } +/** +* clocksource_read: - Read nanosecond delta from clocksource. +* @cs: pointer to clocksource being read +* +* Read from the clock source and return the clock value converted +* to nanoseconds. +*/ +static inline s64 clocksource_read (struct clocksource *cs) +{ +cycle_t cycle_now, cycle_delta;
include/linux/clocksource.h static inline s64 cyc2ns(struct clocksou · line 306
Quoted 10 lines
include/linux/clocksource.h +10
-0+ +/* read clocksource: */ + cycle_now = cs->read (cs); + +/* calculate the delta since the last update_wall_time: */ + cycle_delta = (cycle_now - cs->cycle_last) & cs->mask; + +/* convert to nanoseconds: */ +return cyc2ns (cs, cycle_delta); +}
Oof. So you took out clocksource_read() only to replace with a different function with the same name? If this move is necessary, could we call these clocksource_get_ns()/clocksource_get_raw_ns() to avoid the confusion?
thanks
-john
Re: [RFC][patch 4/5] clocksource_read/clocksource_read_raw inline functions
Quoted 41 lines · On Tue, 21 Jul 2009 15:01:14 -0700 john stultz <[email protected]> wrote:
Quoted 40 lines · Martin Schwidefsky
plain text document attachment (clocksource-read-ns.diff)
From: Martin Schwidefsky <[email protected]>Add
clocksource_read/clocksource_read_rawinline functions and use them for getnstimeofday,ktime_get,ktime_get_tsand getrawmonotonic.Cc Ingo Molnar <[email protected]>Cc Thomas Gleixner <[email protected]>Cc john stultz <[email protected]>Signed-off-by Martin Schwidefsky <[email protected]>--- include/linux/clocksource.h | 45 +++++++++++++++++++++++++++++++++++++++++++ kernel/time/timekeeping.c | 46 +++----------------------------------------- 2 files changed, 49 insertions(+), 42 deletions(-)Index:
linux-2.6/include/linux/clocksource.h
===================================================================include/linux/clocksource.h +17
-0@@ -284,6 +284,51 @@ static inline s64 cyc2ns(struct clocksoureturn ret; } +/** +* clocksource_read: - Read nanosecond delta from clocksource. +* @cs: pointer to clocksource being read +* +* Read from the clock source and return the clock value converted +* to nanoseconds. +*/ +static inline s64 clocksource_read (struct clocksource *cs) +{ +cycle_t cycle_now, cycle_delta; + +/* read clocksource: */ + cycle_now = cs->read (cs); + +/* calculate the delta since the last update_wall_time: */ + cycle_delta = (cycle_now - cs->cycle_last) & cs->mask; +
Quoted 8 lines
include/linux/clocksource.h static inline s64 cyc2ns(struct clocksou · line 306
Quoted 3 lines
include/linux/clocksource.h +3
-0+/* convert to nanoseconds: */ +return cyc2ns (cs, cycle_delta); +}Oof. So you took out
clocksource_read()only to replace with a different function with the same name? If this move is necessary, could we call these clocksource_get_ns()/clocksource_get_raw_ns() to avoid the confusion?
The code that uses the old clocksource_read is accessing other fields from the struct clocksource. So the old inline function seems redundant to me. The new one do more and returns a value where you don't need anything else from the clocksource (namely nanoseconds). To reuse the name adds a little confusion while you look at the patches, the end result should be pretty clear, no? But if you prefer to add the _ns suffix that is fine with me.
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.
RFC PATCH 5/5 update clocksource with stop_machine
In reply to Martin SchwidefskyFrom: Martin Schwidefsky <[email protected]>
update_wall_time calls change_clocksource HZ times per second to check if a new clock source is available. In close to 100% of all calls there is no new clock. Replace the tick based check by an update done with stop_machine.
--- include/linux/clocksource.h | 2 - kernel/time/clocksource.c | 82 ++++++++++++++++---------------------------- kernel/time/timekeeping.c | 2 - 3 files changed, 31 insertions(+), 55 deletions(-)
include/linux/clocksource.h +0 -2
@@ -340,10 +340,8 @@ extern void clocksource_resume(void);#ifdef CONFIG_GENERIC_TIME extern void clocksource_forward_now (void ); -extern void change_clocksource (void );#else static inline void clocksource_forward_now (void ) { } -static inline void change_clocksource (void ) { }#endif #ifdef CONFIG_GENERIC_TIME_VSYSCALL
kernel/time/clocksource.c +31 -51
167 lines not on this page · raw message
kernel/time/timekeeping.c +0 -2
@@ -491,8 +491,6 @@ void update_wall_time(void)update_xtime_cache (cyc2ns (clock, offset)); -/* check to see if there is a new clocksource to use */ -change_clocksource ();update_vsyscall (&xtime, clock); }
Signature · 4 lines
blue skies, Martin. "Reality continues to ruin my life." - Calvin.