Hi!
With the introduction of USER_HZ some counters may wrap much faster on 32 bit
platforms. On i386 HZ=1000 which means that all kinds of time related counters
may wrap within (4294967296 / 1000 / 60 / 60 / 24) = 49 days.
This patch is an attempt to overcome this problem by replacing the relevant
counters with 64 bit counters. This introduces the necessity to do additional
locking when these counters are accessed as operations on these counters no
longer are atomic.
To limit the size of this message the attached patch is restricted to the
kernel_stat.h and sched.h files, just to give an impression. The full patch
can be downloaded from:
ftp://thttp://www.vertis.nl/pub/linux/userhz/linux-2.5.38-u64-1.patch
Rolf Fokkens
[email protected]
diff -ruN linux-2.5.38.orig/include/linux/kernel_stat.h linux-2.5.38/include/linux/kernel_stat.h
--- linux-2.5.38.orig/include/linux/kernel_stat.h Sun Aug 11 03:41:27 2002
+++ linux-2.5.38/include/linux/kernel_stat.h Sun Sep 22 20:28:55 2002
@@ -5,6 +5,8 @@
#include <asm/irq.h>
#include <linux/smp.h>
#include <linux/threads.h>
+#include <linux/time.h>
+#include <linux/spinlock.h>
/*
* 'kernel_stat.h' contains the definitions needed for doing
@@ -16,9 +18,10 @@
#define DK_MAX_DISK 16
struct kernel_stat {
- unsigned int per_cpu_user[NR_CPUS],
- per_cpu_nice[NR_CPUS],
- per_cpu_system[NR_CPUS];
+ rwlock_t times_lock;
+ u64 uptime,
+ cpu_utime[NR_CPUS], cpu_ntime[NR_CPUS],
+ cpu_stime[NR_CPUS];
unsigned int dk_drive[DK_MAX_MAJOR][DK_MAX_DISK];
unsigned int dk_drive_rio[DK_MAX_MAJOR][DK_MAX_DISK];
unsigned int dk_drive_wio[DK_MAX_MAJOR][DK_MAX_DISK];
diff -ruN linux-2.5.38.orig/include/linux/sched.h linux-2.5.38/include/linux/sched.h
--- linux-2.5.38.orig/include/linux/sched.h Sat Sep 21 13:03:22 2002
+++ linux-2.5.38/include/linux/sched.h Sun Sep 22 20:28:39 2002
@@ -82,9 +82,6 @@
load += n*(FIXED_1-exp); \
load >>= FSHIFT;
-#define CT_TO_SECS(x) ((x) / HZ)
-#define CT_TO_USECS(x) (((x) % HZ) * 1000000/HZ)
-
extern int nr_threads;
extern int last_pid;
extern unsigned long nr_running(void);
@@ -155,8 +152,6 @@
extern void cpu_init (void);
extern void trap_init(void);
extern void update_process_times(int user);
-extern void update_one_process(struct task_struct *p, unsigned long user,
- unsigned long system, int cpu);
extern void scheduler_tick(int user_tick, int system);
extern unsigned long cache_decay_ticks;
@@ -340,9 +335,12 @@
unsigned long it_real_value, it_prof_value, it_virt_value;
unsigned long it_real_incr, it_prof_incr, it_virt_incr;
struct timer_list real_timer;
- unsigned long utime, stime, cutime, cstime;
- unsigned long start_time;
- long per_cpu_utime[NR_CPUS], per_cpu_stime[NR_CPUS];
+
+ rwlock_t times_lock;
+ u64 utime, stime, cutime, cstime;
+ u64 cpu_utime[NR_CPUS], cpu_stime[NR_CPUS];
+ u64 start_time;
+
/* mm fault and swap info: this can arguably be seen as either mm-specific or thread-specific */
unsigned long min_flt, maj_flt, nswap, cmin_flt, cmaj_flt, cnswap;
int swappable:1;
On Mon, Sep 23, 2002 at 12:07:22AM +0200, Rolf Fokkens wrote:
@@ -340,9 +335,12 @@
unsigned long it_real_value, it_prof_value, it_virt_value;
unsigned long it_real_incr, it_prof_incr, it_virt_incr;
struct timer_list real_timer;
- unsigned long utime, stime, cutime, cstime;
- unsigned long start_time;
- long per_cpu_utime[NR_CPUS], per_cpu_stime[NR_CPUS];
+
+ rwlock_t times_lock;
+ u64 utime, stime, cutime, cstime;
+ u64 cpu_utime[NR_CPUS], cpu_stime[NR_CPUS];
+ u64 start_time;
+
Hmm. Isn't task_t bloated enough already? I'd rather remove them than
make them 64-bit.
Thanks,
Bill
On Mon, 2002-09-23 at 18:08, Rolf Fokkens wrote:
> On Monday 23 September 2002 04:36, William Lee Irwin III wrote:
>
> > - unsigned long utime, stime, cutime, cstime;
> > - unsigned long start_time;
> > - long per_cpu_utime[NR_CPUS], per_cpu_stime[NR_CPUS];
> >
> > Hmm. Isn't task_t bloated enough already? I'd rather remove them than
> > make them 64-bit.
>
> Since nobody else asks this question:
>
> Do you mean to leave out process statistics?
Yes, I think he does.
Having arrays statically created at NR_CPUS inside the task_struct is
just gross. Especially with NR_CPUS=32. That is 128 bytes each! Now
with your changes, it is 256 bytes each!
Sacrifice them to the gods of bloat.
Robert Love
On Monday 23 September 2002 04:36, William Lee Irwin III wrote:
> - unsigned long utime, stime, cutime, cstime;
> - unsigned long start_time;
> - long per_cpu_utime[NR_CPUS], per_cpu_stime[NR_CPUS];
>
> Hmm. Isn't task_t bloated enough already? I'd rather remove them than
> make them 64-bit.
Since nobody else asks this question:
Do you mean to leave out process statistics?
On Tuesday 24 September 2002 00:13, Robert Love wrote:
> Having arrays statically created at NR_CPUS inside the task_struct is
> just gross. Especially with NR_CPUS=32. That is 128 bytes each! Now
> with your changes, it is 256 bytes each!
I can understand that. However from a user point of view statistics are very
usefull information, but not specifically the per-processor statistics.
I assume you mean to leave out the per-process statistics? Or do you mean to
kmalloc the per-processor statistics when needed - that is: only when
processes are running or maybe when the user has chosen to turn then on (some
sysctl maybe)?
Rolf
On Tuesday 24 September 2002 00:13, Robert Love wrote:
>> Having arrays statically created at NR_CPUS inside the task_struct is
>> just gross. Especially with NR_CPUS=32. That is 128 bytes each! Now
>> with your changes, it is 256 bytes each!
On Tue, Sep 24, 2002 at 09:04:25AM +0200, Rolf Fokkens wrote:
> I can understand that. However from a user point of view statistics are very
> usefull information, but not specifically the per-processor statistics.
> I assume you mean to leave out the per-process statistics? Or do you mean to
> kmalloc the per-processor statistics when needed - that is: only when
> processes are running or maybe when the user has chosen to turn then on (some
> sysctl maybe)?
I'm in favor of removing them entirely.
Cheers,
Bill
On Tue, 24 Sep 2002, Rolf Fokkens wrote:
> On Monday 23 September 2002 04:36, William Lee Irwin III wrote:
> > - unsigned long utime, stime, cutime, cstime;
> > - unsigned long start_time;
> > - long per_cpu_utime[NR_CPUS], per_cpu_stime[NR_CPUS];
> >
> > Hmm. Isn't task_t bloated enough already? I'd rather remove them than
> > make them 64-bit.
>
> Since nobody else asks this question:
>
> Do you mean to leave out process statistics?
We don't need to leave out process statistics completely, but per-CPU
per-process statistics indeed looks like overkill.
Tim
P.S.: Some work with respect to cleaning up interfaces of 32 bit jiffies
has gone into -dj already, but I'm still waiting for the next -dj release
to sync up.
On Tue, 24 Sep 2002, Rolf Fokkens wrote:
>> Since nobody else asks this question:
>> Do you mean to leave out process statistics?
On Tue, Sep 24, 2002 at 01:47:38PM +0200, Tim Schmielau wrote:
> We don't need to leave out process statistics completely, but per-CPU
> per-process statistics indeed looks like overkill.
> Tim
> P.S.: Some work with respect to cleaning up interfaces of 32 bit jiffies
> has gone into -dj already, but I'm still waiting for the next -dj release
> to sync up.
The per-cpu per-process stats are the only ones I suggest removing.
NR_CPUS can get large enough to cause a significant amount of bloat.
Cheers,
Bill