Re: [PATCH] x86/apic/vector: Move pr_warn() outside of vector_lock

From: Thomas Gleixner
Date: Sun Mar 28 2021 - 18:05:17 EST


Waiman,

On Sun, Mar 28 2021 at 15:58, Waiman Long wrote:
> It was found that the following circular locking dependency warning
> could happen in some systems:
>
> [ 218.097878] ======================================================
> [ 218.097879] WARNING: possible circular locking dependency detected
> [ 218.097880] 4.18.0-228.el8.x86_64+debug #1 Not tainted
> [ 218.097881] ------------------------------------------------------
> [ 218.097882] systemd/1 is trying to acquire lock:
> [ 218.097883] ffffffff84c27920 (console_owner){-.-.}, at: console_unlock+0x3fb/0x9f0
> [ 218.097886]
> [ 218.097887] but task is already holding lock:
> [ 218.097888] ffffffff84afca78 (vector_lock){-.-.}, at: x86_vector_activate+0xca/0xab0
> [ 218.097891]
> [ 218.097892] which lock already depends on the new lock.
> :
> [ 218.097966] other info that might help us debug this:
> [ 218.097967]
> [ 218.097967] Chain exists of:
> [ 218.097968] console_oc_lock_class --> vector_lock
> [ 218.097972]
> [ 218.097973] Possible unsafe locking scenario:
> [ 218.097973]
> [ 218.097974] CPU0 CPU1
> [ 218.097975] ---- ----
> [ 218.097975] lock(vector_lock);
> [ 218.097977] lock(&irq_desc_lock_class);
> [ 218.097980] lock(vector_lock);
> [ 218.097981] lock(console_owner);
> [ 218.097983]
> [ 218.097984] *** DEADLOCK ***

can you please post the full lockdep output?

> This lockdep warning was causing by printing of the warning message:
>
> [ 218.095152] irq 3: Affinity broken due to vector space exhaustion.
>
> It looks that this warning message is relatively more common than
> the other warnings in arch/x86/kernel/apic/vector.c. To avoid this
> potential deadlock scenario, this patch moves all the pr_warn() calls
> in the vector.c file outside of the vector_lock critical sections.

Definitely not.

> -static int activate_reserved(struct irq_data *irqd)
> +static int activate_reserved(struct irq_data *irqd, unsigned long flags,
> + bool *unlocked)
> {
> struct apic_chip_data *apicd = apic_chip_data(irqd);
> int ret;
> @@ -410,6 +411,8 @@ static int activate_reserved(struct irq_data *irqd)
> */
> if (!cpumask_subset(irq_data_get_effective_affinity_mask(irqd),
> irq_data_get_affinity_mask(irqd))) {
> + raw_spin_unlock_irqrestore(&vector_lock, flags);
> + *unlocked = true;

What?

> pr_warn("irq %u: Affinity broken due to vector space exhaustion.\n",
> irqd->irq);
> }
> @@ -446,6 +449,7 @@ static int x86_vector_activate(struct irq_domain *dom, struct irq_data *irqd,
> {
> struct apic_chip_data *apicd = apic_chip_data(irqd);
> unsigned long flags;
> + bool unlocked = false;
> int ret = 0;
>
> trace_vector_activate(irqd->irq, apicd->is_managed,
> @@ -459,8 +463,9 @@ static int x86_vector_activate(struct irq_domain *dom, struct irq_data *irqd,
> else if (apicd->is_managed)
> ret = activate_managed(irqd);
> else if (apicd->has_reserved)
> - ret = activate_reserved(irqd);
> - raw_spin_unlock_irqrestore(&vector_lock, flags);
> + ret = activate_reserved(irqd, flags, &unlocked);
> + if (!unlocked)
> + raw_spin_unlock_irqrestore(&vector_lock, flags);

Even moar what?

> return ret;
> }

This turns that code into complete unreadable gunk. No way.

Thanks,

tglx