On Fri, 2014-04-11 at 20:57 +0200, Sebastian Andrzej Siewior wrote:
> This -RT series didn't crashed within ~4h testing on my ARM and
> x86-32.
> x86-64 crashed after I started hackbench. I figured out that the crash
> does not happen with lazy-preempt disabled. Therefore the last but one
> patch in the queue disables lazy preempt on x86-64. With this change the
> test box survived ~2h without a crash. I look at this later but it looks
> good now.
I think the below fixes it (in a more or less minimalist way), but it's
not very pretty. Methinks it would be prettier to either clone the x86
percpu + fold logic, or neutralize that optimization completely when
PREEMPT_LAZY is enabled.
x86_32 bit is completely untested, x86_64 hasn't exploded.. yet :)
---
include/linux/preempt.h | 3 +--
arch/x86/include/asm/preempt.h | 8 ++++++++
arch/x86/kernel/asm-offsets.c | 1 +
arch/x86/kernel/entry_32.S | 9 ++++++---
arch/x86/kernel/entry_64.S | 7 +++++--
5 files changed, 21 insertions(+), 7 deletions(-)
--- a/include/linux/preempt.h
+++ b/include/linux/preempt.h
@@ -126,8 +126,7 @@ do { \
#define preempt_enable_notrace() \
do { \
barrier(); \
- if (unlikely(__preempt_count_dec_and_test() || \
- test_thread_flag(TIF_NEED_RESCHED_LAZY))) \
+ if (unlikely(__preempt_count_dec_and_test())) \
__preempt_schedule_context(); \
} while (0)
#else
--- a/arch/x86/include/asm/preempt.h
+++ b/arch/x86/include/asm/preempt.h
@@ -94,7 +94,11 @@ static __always_inline bool __preempt_co
{
if (____preempt_count_dec_and_test())
return true;
+#ifdef CONFIG_PREEMPT_LAZY
return test_thread_flag(TIF_NEED_RESCHED_LAZY);
+#else
+ return false;
+#endif
}
/*
@@ -102,8 +106,12 @@ static __always_inline bool __preempt_co
*/
static __always_inline bool should_resched(void)
{
+#ifdef CONFIG_PREEMPT_LAZY
return unlikely(!__this_cpu_read_4(__preempt_count) || \
test_thread_flag(TIF_NEED_RESCHED_LAZY));
+#else
+ return unlikely(!__this_cpu_read_4(__preempt_count));
+#endif
}
#ifdef CONFIG_PREEMPT
--- a/arch/x86/kernel/asm-offsets.c
+++ b/arch/x86/kernel/asm-offsets.c
@@ -72,4 +72,5 @@ void common(void) {
BLANK();
DEFINE(PTREGS_SIZE, sizeof(struct pt_regs));
+ DEFINE(_PREEMPT_ENABLED, PREEMPT_ENABLED);
}
--- a/arch/x86/kernel/entry_32.S
+++ b/arch/x86/kernel/entry_32.S
@@ -365,19 +365,22 @@ ENTRY(resume_kernel)
need_resched:
# preempt count == 0 + NEED_RS set?
cmpl $0,PER_CPU_VAR(__preempt_count)
+#ifndef CONFIG_PREEMPT_LAZY
+ jnz restore_all
+#else
jz test_int_off
# atleast preempt count == 0 ?
- cmpl $_TIF_NEED_RESCHED,PER_CPU_VAR(__preempt_count)
+ cmpl $_PREEMPT_ENABLED,PER_CPU_VAR(__preempt_count)
jne restore_all
cmpl $0,TI_preempt_lazy_count(%ebp) # non-zero preempt_lazy_count ?
jnz restore_all
- testl $_TIF_NEED_RESCHED_LAZY, %ecx
+ testl $_TIF_NEED_RESCHED_LAZY, TI_flags(%ebp)
jz restore_all
-
test_int_off:
+#endif
testl $X86_EFLAGS_IF,PT_EFLAGS(%esp) # interrupts off (exception
path) ?
jz restore_all
call preempt_schedule_irq
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -1104,10 +1104,13 @@ ENTRY(native_iret)
/* rcx: threadinfo. interrupts off. */
ENTRY(retint_kernel)
cmpl $0,PER_CPU_VAR(__preempt_count)
+#ifndef CONFIG_PREEMPT_LAZY
+ jnz retint_restore_args
+#else
jz check_int_off
# atleast preempt count == 0 ?
- cmpl $_TIF_NEED_RESCHED,PER_CPU_VAR(__preempt_count)
+ cmpl $_PREEMPT_ENABLED,PER_CPU_VAR(__preempt_count)
jnz retint_restore_args
cmpl $0, TI_preempt_lazy_count(%rcx)
@@ -1115,8 +1118,8 @@ ENTRY(retint_kernel)
bt $TIF_NEED_RESCHED_LAZY,TI_flags(%rcx)
jnc retint_restore_args
-
check_int_off:
+#endif
bt $9,EFLAGS-ARGOFFSET(%rsp) /* interrupts off? */
jnc retint_restore_args
call preempt_schedule_irq
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
