Message ID | 1480427118-5126-14-git-send-email-al.kochet@gmail.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Am Dienstag, 29. November 2016, 16:45:18 schrieb Alexander Kochetkov: > Currently rockchip_timer can be used as a scheduler clock. We properly > marked rk_timer_sched_clock_read() as notrace but we then call another > function rk_timer_counter_read() that _wasn't_ notrace. > > Having a traceable function in the sched_clock() path leads to a recursion > within ftrace and a kernel crash. > > Fix this by adding an extra notrace function to keep other users of > rk_timer_counter_read() traceable. > > Signed-off-by: Alexander Kochetkov <al.kochet@gmail.com> you introduced the issue yourself in patch 11/13. In general any patch should never leave the kernel in a worse state than it was before, so no patch should ever introduce known issues itself. In that line of thought, don't patches 10+11 introduce warnings about unused functions as well when applied without patch 12? Maybe merge them? Heiko > --- > drivers/clocksource/rockchip_timer.c | 9 +++++++-- > 1 file changed, 7 insertions(+), 2 deletions(-) > > diff --git a/drivers/clocksource/rockchip_timer.c > b/drivers/clocksource/rockchip_timer.c index 1af80a0..a127822 100644 > --- a/drivers/clocksource/rockchip_timer.c > +++ b/drivers/clocksource/rockchip_timer.c > @@ -87,7 +87,7 @@ static void rk_timer_update_counter(u64 cycles, struct > rk_timer *timer) writel_relaxed(upper, timer->base + TIMER_LOAD_COUNT1); > } > > -static u64 rk_timer_counter_read(struct rk_timer *timer) > +static u64 notrace _rk_timer_counter_read(struct rk_timer *timer) > { > u64 counter; > u32 lower; > @@ -106,6 +106,11 @@ static u64 rk_timer_counter_read(struct rk_timer > *timer) return counter; > } > > +static u64 rk_timer_counter_read(struct rk_timer *timer) > +{ > + return _rk_timer_counter_read(timer); > +} > + > static void rk_timer_interrupt_clear(struct rk_timer *timer) > { > writel_relaxed(1, timer->base + TIMER_INT_STATUS); > @@ -168,7 +173,7 @@ static u64 notrace rk_timer_sched_clock_read(void) > { > struct rk_clocksource *_cs = &cs_timer; > > - return ~rk_timer_counter_read(&_cs->timer); > + return ~_rk_timer_counter_read(&_cs->timer); > } > > static int __init rk_timer_init(struct device_node *np, u32 ctrl_reg)
Hello Heiko! Thank you for patch review! > 29 нояб. 2016 г., в 18:01, Heiko Stübner <heiko@sntech.de> написал(а): > > you introduced the issue yourself in patch 11/13. In general any patch should > never leave the kernel in a worse state than it was before, so no patch should > ever introduce known issues itself. Agree. > In that line of thought, don't patches 10+11 introduce warnings about unused > functions as well when applied without patch 12? 7 - yes 10 - not 11 - yes > 29 нояб. 2016 г., в 18:03, Heiko Stübner <heiko@sntech.de> написал(а): > > Then why do you need another patch to remove them and don't do that in the > patch removing their respective usage? To make 7 more readable. Overwise patch messed up and unreadable. It’s hard to track what going on in the patch. > Maybe merge them? I'll merge all of them. P.S. here comment from another thead about the patches. http://lists.infradead.org/pipermail/linux-arm-kernel/2016-November/thread.html#470957 > 29 нояб. 2016 г., в 18:07, Robin Murphy <robin.murphy@arm.com> написал(а): > > 3288 (and probably anything newer) is irrelevant to this discussion, as > it has the arch timer interface - that may be busted in other ways (such > as not being correctly set up by firmware and not being always-on as it > should), but frequency is not one of them. This only affects > Cortex-A9/A5 based parts. So I update comments for patches. Looks, like I study archeology with my patches, as all new ARM chips not affected by the problem anymore. Regards, Alexander.
diff --git a/drivers/clocksource/rockchip_timer.c b/drivers/clocksource/rockchip_timer.c index 1af80a0..a127822 100644 --- a/drivers/clocksource/rockchip_timer.c +++ b/drivers/clocksource/rockchip_timer.c @@ -87,7 +87,7 @@ static void rk_timer_update_counter(u64 cycles, struct rk_timer *timer) writel_relaxed(upper, timer->base + TIMER_LOAD_COUNT1); } -static u64 rk_timer_counter_read(struct rk_timer *timer) +static u64 notrace _rk_timer_counter_read(struct rk_timer *timer) { u64 counter; u32 lower; @@ -106,6 +106,11 @@ static u64 rk_timer_counter_read(struct rk_timer *timer) return counter; } +static u64 rk_timer_counter_read(struct rk_timer *timer) +{ + return _rk_timer_counter_read(timer); +} + static void rk_timer_interrupt_clear(struct rk_timer *timer) { writel_relaxed(1, timer->base + TIMER_INT_STATUS); @@ -168,7 +173,7 @@ static u64 notrace rk_timer_sched_clock_read(void) { struct rk_clocksource *_cs = &cs_timer; - return ~rk_timer_counter_read(&_cs->timer); + return ~_rk_timer_counter_read(&_cs->timer); } static int __init rk_timer_init(struct device_node *np, u32 ctrl_reg)
Currently rockchip_timer can be used as a scheduler clock. We properly marked rk_timer_sched_clock_read() as notrace but we then call another function rk_timer_counter_read() that _wasn't_ notrace. Having a traceable function in the sched_clock() path leads to a recursion within ftrace and a kernel crash. Fix this by adding an extra notrace function to keep other users of rk_timer_counter_read() traceable. Signed-off-by: Alexander Kochetkov <al.kochet@gmail.com> --- drivers/clocksource/rockchip_timer.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)