2.6.7-rc2 mismatched preempt count in arch/i386/kernel/irq.c
2.6.7-rc2 (and earlier) arch/i386/kernel/irq.c::do_IRQ() calls irq_exit() which expands into preempt_enable_no_resched(), amongst others. With CONFIG_PREEMPT=y, preempt_enable_no_resched() does dec_preempt_count(). Where is the corresponding inc_preempt_count? AFAICT the use of preempt_enable_no_resched() is unbalanced.
Quoted 5 lines · Keith Owens
2.6.7-rc2 (and earlier) arch/i386/kernel/irq.c::do_IRQ() calls
irq_exit()which expands intopreempt_enable_no_resched(), amongst others. WithCONFIG_PREEMPT=y,preempt_enable_no_resched()doesdec_preempt_count(). Where is the correspondinginc_preempt_count? AFAICT the use ofpreempt_enable_no_resched()is unbalanced.
I believe IRQ_EXIT_OFFSET takes care of that;
#define IRQ_EXIT_OFFSET (HARDIRQ_OFFSET-1)
#define irq_exit() \
do { \
preempt_count() -= IRQ_EXIT_OFFSET; \
if (!in_interrupt() && softirq_pending(smp_processor_id())) \
do_softirq(); \
preempt_enable_no_resched(); \
} while (0)Quoted 12 lines · Zwane Mwaikambo
Quoted 5 lines · Keith Owens
2.6.7-rc2 (and earlier) arch/i386/kernel/irq.c::do_IRQ() calls
irq_exit()which expands intopreempt_enable_no_resched(), amongst others. WithCONFIG_PREEMPT=y,preempt_enable_no_resched()doesdec_preempt_count(). Where is the correspondinginc_preempt_count? AFAICT the use ofpreempt_enable_no_resched()is unbalanced.I believe
IRQ_EXIT_OFFSETtakes care of that;#define
IRQ_EXIT_OFFSET(HARDIRQ_OFFSET-1)
Quoted 7 lines
#define
irq_exit()\do { \ preempt_count() -= IRQ_EXIT_OFFSET; \ if (!in_interrupt() && softirq_pending(smp_processor_id())) \ do_softirq(); \ preempt_enable_no_resched(); \ } while (0)
How ugly ....
Quoted 6 lines · Keith Owens
2.6.7-rc2 (and earlier) arch/i386/kernel/irq.c::do_IRQ() calls
irq_exit()which expands intopreempt_enable_no_resched(), amongst others. WithCONFIG_PREEMPT=y,preempt_enable_no_resched()doesdec_preempt_count(). Where is the correspondinginc_preempt_count? AFAICT the use ofpreempt_enable_no_resched()is unbalanced.
If preempt is enabled, IRQ_EXIT_OFFSET is HARDIRQ_OFFSET-1.
I believe that is where the preempt_enable_no_resched is balanced.