Re: [PATCH] powerpc/entry: Fix double accounting of user time on interrupt entry

[Aboorva Devarajan]

On Wed, 2026-09-02 at 12:37 +0530, Shrikanth Hegde wrote:

Hi Shrikanth,

> Hi Aboorva.
> 
> On 9/2/26 10:36 AM, Aboorva Devarajan wrote:
> > Since the switch to generic entry, an interrupt taken from user mode
> > accounts user time twice: once in arch_interrupt_enter_prepare() and
> > again in arch_enter_from_user_mode(), which irqentry_enter() invokes
> > for the same interrupt:
> > 
> >    arch_interrupt_enter_prepare()
> >        account_cpu_user_entry()
> >    irqentry_enter()
> >      arch_enter_from_user_mode()
> >        account_cpu_user_entry()
> > 
> > account_cpu_user_entry() accumulates the time spent in user mode
> > since the last return to user space, so the second call charges the
> > same interval again.
> > 
> > With CONFIG_VIRT_CPU_ACCOUNTING_NATIVE=y this roughly doubles the
> > reported user time of any workload that takes interrupts. On a
> > pseries LPAR, ps/top show ~200% CPU for a single-threaded CPU-bound
> > loop, and time(1) reports user time about twice the elapsed time.
> > 
> 
> Could you please run mpstat with 50% or less loading workload like stress-ng
> and document the difference in changelog.

Sure, will add it in v2.

1. On the same pseries LPAR, taskset -c 6 stress-ng --cpu 1 --cpu-load 50:

Without the patch, the reported %usr is inflated, while with the patch
it is close to the expected 50%.

Without Patch:
06:43:09 AM  CPU    %usr**   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
06:43:10 AM    6   69.74     0.00    0.00    0.00    0.00    0.00    0.00    0.00    0.00   30.26

With Patch:
06:29:33 AM  CPU    %usr**   %nice    %sys %iowait    %irq   %soft  %steal  %guest  %gnice   %idle
06:29:34 AM    6   48.51     0.00    0.00    0.00    0.99    0.00    0.00    0.00    0.00   50.50

2. Single-threaded CPU-bound workload:

taskset -c 100 python3 -c 'while True: pass' (executed for ~3 seconds)

Without patch:
TIME         ELAPSED %CPU
00:00:06       00:02  209

With patch:
TIME         ELAPSED %CPU
00:00:03       00:03  104

> 
> > Remove the accounting from arch_interrupt_enter_prepare() and rely on
> > arch_enter_from_user_mode(), which runs for both syscalls and
> > interrupts. The duplicate account_stolen_time() call is removed the
> > same way.
> > 
> > Fixes: bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
> > Signed-off-by: Aboorva Devarajan <aboorvad@xxxxxxxxxxxxx>
> > ---
> > Verified on a pseries LPAR (CONFIG_VIRT_CPU_ACCOUNTING_NATIVE=y),
> 
> Usual default is VIRT_CPU_ACCOUNTING_GEN, selected by NO_HZ_FULL.
> Most distros usually enable NO_HZ_FULL=y.
> 
> At hindsight, I don't see the issue dependency on it. But better to be sure.

It is only observed with CONFIG_VIRT_CPU_ACCOUNTING_NATIVE.

account_cpu_user_entry() updates utime only with NATIVE accounting.
With CONFIG_VIRT_CPU_ACCOUNTING_GEN or tick-based accounting, it is
an empty stub and therefore a no-op.

So kernels using CONFIG_VIRT_CPU_ACCOUNTING_GEN or tick-based accounting
are not affected.

> > 7.3.0-rc1, single-threaded CPU-bound loop:
> > 
> > Before:
> > 
> >    $ python3 -c 'while True: pass' &
> >    $ sleep 3; ps -p $! -o pid,etime,time,pcpu
> >        PID     ELAPSED     TIME %CPU
> >       4980       00:03 00:00:06  210
> > 
> > After:
> > 
> >    $ python3 -c 'while True: pass' &
> >    $ sleep 3; ps -p $! -o pid,etime,time,pcpu
> >        PID     ELAPSED     TIME %CPU
> >       4951       00:03 00:00:03  105
> > 
> >   arch/powerpc/include/asm/entry-common.h | 6 ++++--
> >   1 file changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/powerpc/include/asm/entry-common.h b/arch/powerpc/include/asm/entry-common.h
> > index c5adb5006361..984294e62568 100644
> > --- a/arch/powerpc/include/asm/entry-common.h
> > +++ b/arch/powerpc/include/asm/entry-common.h
> > @@ -222,8 +222,6 @@ static inline void arch_interrupt_enter_prepare(struct pt_regs *regs)
> >   
> >   	if (user_mode(regs)) {
> >   		kuap_lock();
> > -		account_cpu_user_entry();
> > -		account_stolen_time();
> >   	} else {
> >   		kuap_save_and_lock(regs);
> >   		/*
> > @@ -426,6 +424,10 @@ static __always_inline void arch_enter_from_user_mode(struct pt_regs *regs)
> >   #endif
> >   	kuap_assert_locked();
> >   	booke_restore_dbcr0();
> > +	/*
> > +	 * User and stolen time is accounted here for every entry from
> > +	 * user mode. The interrupt prepare hooks must not account again.
> > +	 */
> >   	account_cpu_user_entry();
> >   	account_stolen_time();
> >   
> > 
> > base-commit: fb442a6673ff1046bf67754957d95880fdb394b5

Thanks,
Aboorva