If we get no user time and no system time allocated since the last account_system_vtime, the system to user time ratio estimate can end up dividing by zero. This was causing a problem noticed by Balbir Singh. Signed-off-by: Michael Neuling <mikey@neuling.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
These don't need to be seen by everyone on every boot. Signed-off-by: Tony Breeds <tony@bakeyournoodle.com> Signed-off-by: Paul Mackerras <paulus@samba.org>
* git://git.kernel.org/pub/scm/linux/kernel/git/mingo/linux-2.6-sched: sched: proper prototype for kernel/sched.c:migration_init() sched: avoid large irq-latencies in smp-balancing sched: fix copy_namespace() <-> sched_fork() dependency in do_fork sched: clean up the wakeup preempt check, #2 sched: clean up the wakeup preempt check sched: wakeup preemption fix sched: remove PREEMPT_RESTRICT sched: turn off PREEMPT_RESTRICT KVM: fix !SMP build error x86: make nmi_cpu_busy() always defined x86: make ipi_handler() always defined sched: cleanup, use NSEC_PER_MSEC and NSEC_PER_SEC sched: reintroduce SMP tunings again sched: restore deterministic CPU accounting on powerpc sched: fix delay accounting regression sched: reintroduce the sched_min_granularity tunable sched: documentation: place_entity() comments sched: fix vslice
Since powerpc started using CONFIG_GENERIC_CLOCKEVENTS, the deterministic CPU accounting (CONFIG_VIRT_CPU_ACCOUNTING) has been broken on powerpc, because we end up counting user time twice: once in timer_interrupt() and once in update_process_times(). This fixes the problem by pulling the code in update_process_times that updates utime and stime into a separate function called account_process_tick. If CONFIG_VIRT_CPU_ACCOUNTING is not defined, there is a version of account_process_tick in kernel/timer.c that simply accounts a whole tick to either utime or stime as before. If CONFIG_VIRT_CPU_ACCOUNTING is defined, then arch code gets to implement account_process_tick. This also lets us simplify the s390 code a bit; it means that the s390 timer interrupt can now call update_process_times even when CONFIG_VIRT_CPU_ACCOUNTING is turned on, and can just implement a suitable account_process_tick(). account_process_tick() now takes the task_struct * as an argument. Tested both with and without CONFIG_VIRT_CPU_ACCOUNTING. Signed-off-by: Paul Mackerras <paulus@samba.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
The decrementer in Book E and 4xx processors interrupts on the transition from 1 to 0, rather than on the 0 to -1 transition as on 64-bit server and 32-bit "classic" (6xx/7xx/7xxx) processors. At the moment we subtract 1 from the count of how many decrementer ticks are required before the next interrupt before putting it into the decrementer, which is correct for server/classic processors, but could possibly cause the interrupt to happen too early on Book E and 4xx if the timebase/decrementer frequency is low. This fixes the problem by making set_dec subtract 1 from the count for server and classic processors, instead of having the callers subtract 1. Since set_dec already had a bunch of ifdefs to handle different processor types, there is no net increase in ugliness. :) Note that calling set_dec(0) may not generate an interrupt on some processors. To make sure that decrementer_set_next_event always calls set_dec with an interval of at least 1 tick, we set min_delta_ns of the decrementer_clockevent to correspond to 2 ticks (2 rather than 1 to compensate for truncations in the conversions between ticks and ns). This also removes a redundant call to set the decrementer to 0x7fffffff - it was already set to that earlier in timer_interrupt. Signed-off-by: Paul Mackerras <paulus@samba.org>
This adds POWERPC specific hooks for scaled time accounting.
POWER6 includes a SPURR register. The SPURR is based off the PURR register
but is scaled based on CPU frequency and issue rates. This gives a more
accurate account of the instructions used per task. The PURR and timebase
will be constant relative to the wall clock, irrespective of the CPU
frequency.
This implementation reads the SPURR register in account_system_vtime which
is only call called on context witch and hard and soft irq entry and exit.
The percentage of user and system time is then estimated using the ratio of
these accounted by the PURR. If the SPURR is not present, the PURR read.
An earlier implementation of this patch read the SPURR whenever the PURR
was read, which included the system call entry and exit path.
Unfortunately this showed a performance regression on lmbench runs, so was
re-implemented.
I've included the lmbench results here when run bare metal on POWER6. 1st
column is the unpatch results. 2nd column is the results using the below
patch and the 3rd is the % diff of these results from the base. 4th and
5th columns are the results and % differnce from the base using the older
patch (SPURR read in syscall entry/exit path).
Base Scaled-Acct SPURR-in-syscall
Result Result % diff Result % diff
Simple syscall: 0.3086 0.3086 0.0000 0.3452 11.8600
Simple read: 0.4591 0.4671 1.7425 0.5044 9.86713
Simple write: 0.4364 0.4366 0.0458 0.4731 8.40971
Simple stat: 2.0055 2.0295 1.1967 2.0669 3.06158
Simple fstat: 0.5962 0.5876 -1.442 0.6368 6.80979
Simple open/close: 3.1283 3.1009 -0.875 3.2088 2.57328
Select on 10 fd's: 0.8554 0.8457 -1.133 0.8667 1.32101
Select on 100 fd's: 3.5292 3.6329 2.9383 3.6664 3.88756
Select on 250 fd's: 7.9097 8.1881 3.5197 8.2242 3.97613
Select on 500 fd's: 15.2659 15.836 3.7357 15.873 3.97814
Select on 10 tcp fd's: 0.9576 0.9416 -1.670 0.9752 1.83792
Select on 100 tcp fd's: 7.248 7.2254 -0.311 7.2685 0.28283
Select on 250 tcp fd's: 17.7742 17.707 -0.375 17.749 -0.1406
Select on 500 tcp fd's: 35.4258 35.25 -0.496 35.286 -0.3929
Signal handler installation: 0.6131 0.6075 -0.913 0.647 5.52927
Signal handler overhead: 2.0919 2.1078 0.7600 2.1831 4.35967
Protection fault: 0.7345 0.7478 1.8107 0.8031 9.33968
Pipe latency: 33.006 16.398 -50.31 33.475 1.42368
AF_UNIX sock stream latency: 14.5093 30.910 113.03 30.715 111.692
Process fork+exit: 219.8 222.8 1.3648 229.37 4.35623
Process fork+execve: 876.14 873.28 -0.32 868.66 -0.8533
Process fork+/bin/sh -c: 2830 2876.5 1.6431 2958 4.52296
File /var/tmp/XXX write bw: 1193497 1195536 0.1708 118657 -0.5799
Pagefaults on /var/tmp/XXX: 3.1272 3.2117 2.7020 3.2521 3.99398
Also, kernel compile times show no difference with this patch applied.
[pbadari@us.ibm.com: Avoid unnecessary PURR reading]
Signed-off-by: Michael Neuling <mikey@neuling.org>
Cc: Balbir Singh <balbir@in.ibm.com>
Cc: Jay Lan <jlan@engr.sgi.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The clockevent bootup message only needs to be KERN_INFO. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Paul Mackerras <paulus@samba.org>
In testing the new clocksource and clockevent code on a PPC601 processor, I discovered that the clockevent multiplier value for the decrementer clockevent was overflowing. Because the RTCL register in the 601 effectively counts at 1GHz (it doesn't actually, but it increases by 128 every 128ns), and the shift value was 32, that meant the multiplier value had to be 2^32, which won't fit in an unsigned long on 32-bit. The same problem would arise on any platform where the timebase frequency was 1GHz or more (not that we actually have any such machines today). This fixes it by reducing the shift value to 16. Doing the calculations with a resolution of 2^-16 nanoseconds (15 femtoseconds) should be quite adequate. :) Signed-off-by: Paul Mackerras <paulus@samba.org>
On old powermacs, we sometimes set the decrementer to 1 in order to trigger a decrementer interrupt, which we use to handle an interrupt that was pending at the time when it was re-enabled. This was causing the decrementer clock event device to call the event function for the next event early, which was causing problems when high-res timers were not enabled. This fixes the problem by recording the timebase value at which the next event should occur, and checking the current timebase against the recorded value in timer_interrupt. If it isn't time for the next event, it just reprograms the decrementer and returns. This also subtracts 1 from the value stored into the decrementer, which is appropriate because the decrementer interrupts on the transition from 0 to -1, not when the decrementer reaches 0. Signed-off-by: Paul Mackerras <paulus@samba.org>
This registers a clock event structure for the decrementer and turns on CONFIG_GENERIC_CLOCKEVENTS, which means that we now don't need most of timer_interrupt(), since the work is done in generic code. For secondary CPUs, their decrementer clockevent is registered when the CPU comes up (the generic code automatically removes the clockevent when the CPU goes down). Signed-off-by: Tony Breeds <tony@bakeyournoodle.com> Signed-off-by: Paul Mackerras <paulus@samba.org>