Aboorva Devarajan <[email protected]> writes:

> Since the switch to generic entry, an interrupt 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()              /* first */
>       irqentry_enter()
>         arch_enter_from_user_mode()
>           account_cpu_user_entry()            /* second */
>
> The second call charges the same interval again, because
> account_cpu_user_entry() accumulates the time spent in user mode
> since the last return to user space.
>
> The two calls come from the GENERIC_ENTRY preparation series,
> where each step was a no-op on its own. Commit 09a9d3a8499d
> ("powerpc: introduce arch_enter_from_user_mode") added the hook
> with the user-time accounting in it, but nothing called it yet.
> Commit 893082ac769b ("powerpc: Prepare for IRQ entry exit")
> copied interrupt_enter_prepare() verbatim into entry-common.h as
> arch_interrupt_enter_prepare(); that copy was equally unused, as
> handlers still called interrupt_enter_prepare().
>
> Commit bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
> made both live. On the syscall side it did the full conversion:
> system_call_exception() now accounts once through the hook via
> syscall_enter_from_user_mode(), rather than calling
> account_cpu_user_entry() directly. On the interrupt side it
> switched the handler macros to arch_interrupt_enter_prepare()
> followed by irqentry_enter(), which also runs the hook, but the
> accounting in arch_interrupt_enter_prepare() was not removed to
> match. The double accounting starts with that commit.
>
> With CONFIG_VIRT_CPU_ACCOUNTING_NATIVE=y this roughly doubles the
> reported user time of any workload that takes interrupts. The
> other accounting modes compile account_cpu_user_entry() to an
> empty stub, so they are not affected.
>
> Remove the accounting from arch_interrupt_enter_prepare() and rely
> on arch_enter_from_user_mode(), which already runs for both
> syscalls and interrupts. The duplicate account_stolen_time() call
> is removed the same way.
>
> On a pseries LPAR a busy loop reports 6s user time in 3s elapsed
> (~210% CPU) before the fix, and 3s (~105% CPU) after it:
>
>   $ python3 -c 'while True: pass' &
>   $ sleep 3; ps -p $! -o etime,time,pcpu
>
>             ELAPSED     TIME  %CPU
>   Before      00:03 00:00:06   210
>   After       00:03 00:00:03   105
>
> A 50% load reports ~70% usr / 30% idle before the fix, and
> ~49% usr / 51% idle after it:
>
>   $ taskset -c 6 stress-ng --cpu 1 --cpu-load 50 &
>   $ mpstat -P 6 1
>
>             CPU    %usr   %idle
>   Before      6   69.74   30.26
>   After       6   48.51   50.50
>
> Fixes: bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
> Reviewed-by: Mukesh Kumar Chaurasiya (IBM) <[email protected]>
> Signed-off-by: Aboorva Devarajan <[email protected]>
> ---

Looks good. Thanks for the descritive commit msg!

Reviewed-by: Ritesh Harjani (IBM) <[email protected]>

Reply via email to