Messages in this thread | | | From | Sven Schnelle <> | Subject | Re: [PATCHv2 0/5] arm64/irqentry: remove duplicate housekeeping of | Date | Tue, 28 Sep 2021 11:52:51 +0200 |
| |
Mark Rutland <mark.rutland@arm.com> writes:
> On Mon, Sep 27, 2021 at 05:09:22PM -0700, Paul E. McKenney wrote: >> On Mon, Sep 27, 2021 at 10:23:18AM +0100, Mark Rutland wrote: >> > On Fri, Sep 24, 2021 at 03:59:54PM -0700, Paul E. McKenney wrote: >> > > On Fri, Sep 24, 2021 at 06:36:15PM +0100, Mark Rutland wrote: >> > > > [Adding Paul for RCU, s390 folk for entry code RCU semantics] >> > > > >> > > > On Fri, Sep 24, 2021 at 09:28:32PM +0800, Pingfan Liu wrote: >> > > > > After introducing arm64/kernel/entry_common.c which is akin to >> > > > > kernel/entry/common.c , the housekeeping of rcu/trace are done twice as >> > > > > the following: >> > > > > enter_from_kernel_mode()->rcu_irq_enter(). >> > > > > And >> > > > > gic_handle_irq()->...->handle_domain_irq()->irq_enter()->rcu_irq_enter() >> > > > > >> > > > > Besides redundance, based on code analysis, the redundance also raise >> > > > > some mistake, e.g. rcu_data->dynticks_nmi_nesting inc 2, which causes >> > > > > rcu_is_cpu_rrupt_from_idle() unexpected. >> > > > >> > > > Hmmm... >> > > > >> > > > The fundamental questionss are: >> > > > >> > > > 1) Who is supposed to be responsible for doing the rcu entry/exit? >> > > > >> > > > 2) Is it supposed to matter if this happens multiple times? >> > > > >> > > > For (1), I'd generally expect that this is supposed to happen in the >> > > > arch/common entry code, since that itself (or the irqchip driver) could >> > > > depend on RCU, and if that's the case thatn handle_domain_irq() >> > > > shouldn't need to call rcu_irq_enter(). That would be consistent with >> > > > the way we handle all other exceptions. >> > > > >> > > > For (2) I don't know whether the level of nesting is suppoosed to >> > > > matter. I was under the impression it wasn't meant to matter in general, >> > > > so I'm a little surprised that rcu_is_cpu_rrupt_from_idle() depends on a >> > > > specific level of nesting. >> > > > >> > > > >From a glance it looks like this would cause rcu_sched_clock_irq() to >> > > > skip setting TIF_NEED_RESCHED, and to not call invoke_rcu_core(), which >> > > > doesn't sound right, at least... >> > > > >> > > > Thomas, Paul, thoughts? >> > > >> > > It is absolutely required that rcu_irq_enter() and rcu_irq_exit() calls >> > > be balanced. Normally, this is taken care of by the fact that irq_enter() >> > > invokes rcu_irq_enter() and irq_exit() invokes rcu_irq_exit(). Similarly, >> > > nmi_enter() invokes rcu_nmi_enter() and nmi_exit() invokes rcu_nmi_exit(). >> > >> > Sure; I didn't mean to suggest those weren't balanced! The problem here >> > is *nesting*. Due to the structure of our entry code and the core IRQ >> > code, when handling an IRQ we have a sequence: >> > >> > irq_enter() // arch code >> > irq_enter() // irq code >> > >> > < irq handler here > >> > >> > irq_exit() // irq code >> > irq_exit() // arch code >> > >> > ... and if we use something like rcu_is_cpu_rrupt_from_idle() in the >> > middle (e.g. as part of rcu_sched_clock_irq()), this will not give the >> > expected result because of the additional nesting, since >> > rcu_is_cpu_rrupt_from_idle() seems to expect that dynticks_nmi_nesting >> > is only incremented once per exception entry, when it does: >> > >> > /* Are we at first interrupt nesting level? */ >> > nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting); >> > if (nesting > 1) >> > return false; >> > >> > What I'm trying to figure out is whether that expectation is legitimate, >> > and assuming so, where the entry/exit should happen. >> >> Oooh... >> >> The penalty for fooling rcu_is_cpu_rrupt_from_idle() is that RCU will >> be unable to detect a userspace quiescent state for a non-nohz_full >> CPU. That could result in RCU CPU stall warnings if a user task runs >> continuously on a given CPU for more than 21 seconds (60 seconds in >> some distros). And this can easily happen if the user has a CPU-bound >> thread that is the only runnable task on that CPU. >> >> So, yes, this does need some sort of resolution. >> >> The traditional approach is (as you surmise) to have only a single call >> to irq_enter() on exception entry and only a single call to irq_exit() >> on exception exit. If this is feasible, it is highly recommended. > > Cool; that's roughly what I was expecting / hoping to hear! > >> In theory, we could have that "1" in "nesting > 1" be a constant supplied >> by the architecture (you would want "3" if I remember correctly) but >> in practice could we please avoid this? For one thing, if there is >> some other path into the kernel for your architecture that does only a >> single irq_enter(), then rcu_is_cpu_rrupt_from_idle() just doesn't stand >> a chance. It would need to compare against a different value depending >> on what exception showed up. Even if that cannot happen, it would be >> better if your architecture could remain in blissful ignorance of the >> colorful details of ->dynticks_nmi_nesting manipulations. > > I completely agree. I think it's much harder to keep that in check than > to enforce a "once per architectural exception" policy in the arch code. > >> Another approach would be for the arch code to supply RCU a function that >> it calls. If there is such a function (or perhaps better, if some new >> Kconfig option is enabled), RCU invokes it. Otherwise, it compares to >> "1" as it does now. But you break it, you buy it! ;-) > > I guess we could look at the exception regs and inspect the original > context, but it sounds overkill... > > I think the cleanest thing is to leave this to arch code, and have the > common IRQ code stay well clear. Unfortunately most architectures > (including arch/arm) still need the common IRQ code to handle this, so > we'll have to make that conditional on Kconfig, something like the below > (build+boot tested only). > > If there are no objections, I'll go check who else needs the same > treatment (IIUC at least s390 will), and spin that as a real > patch/series.
Hmm, s390 doesn't use handle_domain_irq() and doesn't have HANDLE_DOMAIN_IRQ set. So i don't think the patch below applies to s390. However, i'll follow the code to make sure we're not calling irq_enter/irq_exit twice.
> Thanks, > Mark. > > ---->8---- > diff --git a/arch/Kconfig b/arch/Kconfig > index 8df1c7102643..c59475e50e4c 100644 > --- a/arch/Kconfig > +++ b/arch/Kconfig > @@ -225,6 +225,12 @@ config GENERIC_SMP_IDLE_THREAD > config GENERIC_IDLE_POLL_SETUP > bool > > +config ARCH_ENTERS_IRQ > + bool > + help > + An architecture should select this when it performs irq entry > + management itself (e.g. calling irq_enter() and irq_exit()). > + > config ARCH_HAS_FORTIFY_SOURCE > bool > help > diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig > index 5c7ae4c3954b..fa6476bf2b4d 100644 > --- a/arch/arm64/Kconfig > +++ b/arch/arm64/Kconfig > @@ -16,6 +16,7 @@ config ARM64 > select ARCH_ENABLE_MEMORY_HOTREMOVE > select ARCH_ENABLE_SPLIT_PMD_PTLOCK if PGTABLE_LEVELS > 2 > select ARCH_ENABLE_THP_MIGRATION if TRANSPARENT_HUGEPAGE > + select ARCH_ENTERS_IRQ > select ARCH_HAS_CACHE_LINE_SIZE > select ARCH_HAS_DEBUG_VIRTUAL > select ARCH_HAS_DEBUG_VM_PGTABLE > diff --git a/kernel/irq/irqdesc.c b/kernel/irq/irqdesc.c > index 4e3c29bb603c..6affa12222e0 100644 > --- a/kernel/irq/irqdesc.c > +++ b/kernel/irq/irqdesc.c > @@ -677,6 +677,15 @@ int generic_handle_domain_irq(struct irq_domain *domain, unsigned int hwirq) > EXPORT_SYMBOL_GPL(generic_handle_domain_irq); > > #ifdef CONFIG_HANDLE_DOMAIN_IRQ > + > +#ifdef ARCH_ENTERS_IRQ > +#define handle_irq_enter() > +#define handle_irq_exit() > +#else > +#define handle_irq_enter() irq_enter() > +#define handle_irq_exit() irq_exit() > +#endif > + > /** > * handle_domain_irq - Invoke the handler for a HW irq belonging to a domain, > * usually for a root interrupt controller > @@ -693,7 +702,7 @@ int handle_domain_irq(struct irq_domain *domain, > struct irq_desc *desc; > int ret = 0; > > - irq_enter(); > + handle_irq_enter(); > > /* The irqdomain code provides boundary checks */ > desc = irq_resolve_mapping(domain, hwirq); > @@ -702,7 +711,7 @@ int handle_domain_irq(struct irq_domain *domain, > else > ret = -EINVAL; > > - irq_exit(); > + handle_irq_exit(); > set_irq_regs(old_regs); > return ret; > }
| |