A while back (before [1]), i386 had this in the tail of its irq
handling code:
need_resched:
movl TI_flags(%ebp), %ecx # need_resched set ?
testb $_TIF_NEED_RESCHED, %cl
jz restore_all
testl $IF_MASK,EFLAGS(%esp) # interrupts off (exception path) ?
jz restore_all
sti
call preempt_schedule
cli
movl $0,TI_preempt_count(%ebp)
jmp need_resched
preempt_schedule() already had an inner need_resched() loop introduced by
commit 4d0b85ea4610 ("[PATCH] kernel preemption bits (2/2)")
and the outer loop was needed since, as the above commit explains, it
was possible to return from preempt_schedule(), get an interrupt and
have need_resched() true.
This was eventually changed when preempt_schedule_irq() was introduced
by [1]:
commit b268264c6299 ("[PATCH] sched: fix preemption race (Core/i386)")
which moved the enabling & disabling of interrupts inside the then
new preempt_schedule_irq(). From then on, the arch-code loop was no
longer necessary, providing preempt_schedule_irq() was used. This was
talked over on LKML in [2].
It's worth noting that it seems most archs calling
preempt_schedule_irq() have that outer loop, and for most of them it
really looks like they could get rid of it as well.
FWIW the suspects are
$ grep -r -I "preempt_schedule_irq" arch/ | cut -d/ -f2 | sort | uniq
arc
arm
arm64
c6x
csky
h8300
ia64
m68k
microblaze
mips
nds32
nios2
parisc
powerpc
s390
sh
sparc
x86
xtensa
- Patches 1-2 remove the loop for arm & arm64
- Patch 3 adds a bit of documentation to point out the loop isn't needed
to try and spread the word.
[2]: https://lore.kernel.org/lkml/cc989920-a13b-d53b-db83-1584a7f53edc@xxxxxxx/
Valentin Schneider (3):
arm64: entry: Remove unneeded need_resched() loop
ARM: entry: Remove unneeded need_resched() loop
sched/Documentation: Point out use of preempt_schedule_irq()
Documentation/scheduler/sched-arch.txt | 10 ++++++++++
arch/arm/kernel/entry-armv.S | 12 +-----------
arch/arm64/kernel/entry.S | 11 +----------
3 files changed, 12 insertions(+), 21 deletions(-)
--
2.20.1
