There are two issues with the current layout of tick_irq_enter() as
it's called before HARDIRQ_OFFSET is incremented:
1) It's not correctly handled by lockdep which doesn't consider it as
hardirq context. And jiffies/timekeeping update take a few interesting
locks.
2) Softirqs need to be explicitly disabled around it to prevent ksoftirqd
from being spuriously woken up.
The current call dependency prevents tick_irq_enter() from being moved
after HARDIRQ_OFFSET: account_irq_enter_time() needs to be called before
HARDIRQ_OFFSET incrementation due to cputime index dispatch and it must
be called after tick_irq_enter() which updates the clocks that may be
necessary for cputime accounting.
Here is a proposal to fix this layout.
(The EXPORT_SYMBOL_GPL() in vtime will likely disappear in the next take
as they don't seem to be necessary anymore, but I'll need to check
that thoroughly).
git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
irq/core
HEAD: 9502ee20aed8bb847176e1d7d83ccd0625430744
Frederic Weisbecker (4):
sched/vtime: Consolidate IRQ time accounting
s390/vtime: Convert to consolidated IRQ time accounting
irqtime: Move irqtime entry accounting after irq offset incrementation
irq: Call tick_irq_enter() inside HARDIRQ_OFFSET
arch/ia64/kernel/time.c | 22 +++++++---
arch/powerpc/kernel/time.c | 60 ++++++++++++++++++++--------
arch/s390/include/asm/vtime.h | 1 -
arch/s390/kernel/vtime.c | 60 ++++++++++++++++++----------
include/linux/hardirq.h | 4 +-
include/linux/vtime.h | 18 ++++-----
kernel/sched/cputime.c | 75 +++++++++++++++++++++++++++--------
kernel/softirq.c | 16 +++-----
8 files changed, 176 insertions(+), 80 deletions(-)
--
2.25.1
