[PATCH v2 07/17] irqchip/gic: Atomically update affinity

[Marc Zyngier]

Hi Geert,

On Thu, 09 Sep 2021 16:22:01 +0100,
Geert Uytterhoeven <geert at linux-m68k.org> wrote:
> 
> Hi Marc, Russell,
> 
> On Wed, Jun 24, 2020 at 9:59 PM Marc Zyngier <maz at kernel.org> wrote:
> > The GIC driver uses a RMW sequence to update the affinity, and
> > relies on the gic_lock_irqsave/gic_unlock_irqrestore sequences
> > to update it atomically.
> >
> > But these sequences only expend into anything meaningful if
> > the BL_SWITCHER option is selected, which almost never happens.
> >
> > It also turns out that using a RMW and locks is just as silly,
> > as the GIC distributor supports byte accesses for the GICD_TARGETRn
> > registers, which when used make the update atomic by definition.
> >
> > Drop the terminally broken code and replace it by a byte write.
> >
> > Fixes: 04c8b0f82c7d ("irqchip/gic: Make locking a BL_SWITCHER only feature")
> > Cc: stable at vger.kernel.org
> > Signed-off-by: Marc Zyngier <maz at kernel.org>
> 
> Thanks for your patch, which is now commit 005c34ae4b44f085
> ("irqchip/gic: Atomically update affinity"), to which I bisected a hard
> lock-up during boot on the Renesas EMMA Mobile EV2-based KZM-A9-Dual
> board, which has a dual Cortex-A9 with PL390.
> 
> Despite the ARM Generic Interrupt Controller Architecture Specification
> (both version 1.0 and 2.0) stating that the Interrupt Processor Targets
> Registers are byte-accessible, the EMMA Mobile EV2 User's Manual
> states that the interrupt registers can be accessed via the APB bus,
> in 32-bit units.  Using byte accesses locks up the system.

Urgh. That is definitely a pretty poor integration. How about the
priority registers? I guess they suffer from the same issue...

> Unfortunately I only have remote access to the board showing the
> issue.  I did check that adding the writeb_relaxed() before the
> writel_relaxed() that was used before also causes a lock-up, so the
> issue is not an endian mismatch.
> Looking at the driver history, these registers have always been
> accessed using 32-bit accesses before.  As byte accesses lead
> indeed to simpler code, I'm wondering if they had been tried before,
> and caused issues before?

Not that I know. A lock was probably fine on a two CPU system. Less so
on a busy 8 CPU machine where interrupts are often migrated. The GIC
architecture makes a point in not requiring locking for most of the
registers that can be accessed concurrently.

> Since you said the locking was bogus before, due to the reliance on
> the BL_SWITCHER option, I'm not suggesting a plain revert, but I'm
> wondering what kind of locking you suggest to use instead?

There isn't much we can do aside from reintroducing the RMW+spinlock
approach, and for real this time. It would have to be handled as a
quirk though, as I'm not keen on reintroducing this for all systems.

I wrote the patchlet below, which is totally untested. Please give it
a go and let me know if it helps.

	M.

diff --git a/drivers/irqchip/irq-gic.c b/drivers/irqchip/irq-gic.c
index d329ec3d64d8..dca40a974b7a 100644
--- a/drivers/irqchip/irq-gic.c
+++ b/drivers/irqchip/irq-gic.c
@@ -107,6 +107,8 @@ static DEFINE_RAW_SPINLOCK(cpu_map_lock);
 
 #endif
 
+static DEFINE_STATIC_KEY_FALSE(needs_rmw_access);
+
 /*
  * The GIC mapping of CPU interfaces does not necessarily match
  * the logical CPU numbering.  Let's use a mapping as returned
@@ -774,6 +776,25 @@ static int gic_pm_init(struct gic_chip_data *gic)
 #endif
 
 #ifdef CONFIG_SMP
+static void rmw_writeb(u8 bval, void __iomem *addr)
+{
+	static DEFINE_RAW_SPINLOCK(rmw_lock);
+	unsigned long offset = (unsigned long)addr & ~3UL;
+	unsigned long shift = offset * 8;
+	unsigned long flags;
+	u32 val;
+
+	raw_spin_lock_irqsave(&rmw_lock, flags);
+
+	addr -= offset;
+	val = readl_relaxed(addr);
+	val &= ~(0xffUL << shift);
+	val |= (u32)bval << shift;
+	writel_relaxed(val, addr);
+
+	raw_spin_unlock_irqrestore(&rmw_lock, flags);
+}
+
 static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
 			    bool force)
 {
@@ -788,7 +809,10 @@ static int gic_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
 	if (cpu >= NR_GIC_CPU_IF || cpu >= nr_cpu_ids)
 		return -EINVAL;
 
-	writeb_relaxed(gic_cpu_map[cpu], reg);
+	if (static_branch_unlikely(&needs_rmw_access))
+		rmw_writeb(gic_cpu_map[cpu], reg);
+	else
+		writeb_relaxed(gic_cpu_map[cpu], reg);
 	irq_data_update_effective_affinity(d, cpumask_of(cpu));
 
 	return IRQ_SET_MASK_OK_DONE;
@@ -1375,6 +1399,29 @@ static bool gic_check_eoimode(struct device_node *node, void __iomem **base)
 	return true;
 }
 
+static bool gic_enable_rmw_access(void *data)
+{
+	/*
+	 * The EMEV2 class of machines has a broken interconnect, and
+	 * locks up on accesses that are less than 32bit. So far, only
+	 * the affinity setting requires it.
+	 */
+	if (of_machine_is_compatible("renesas,emev2")) {
+		static_branch_enable(&needs_rmw_access);
+		return true;
+	}
+
+	return false;
+}
+
+static const struct gic_quirk gic_quirks[] = {
+	{
+		.desc		= "Implementation with broken byte access",
+		.compatible	= "arm,pl390",
+		.init		= gic_enable_rmw_access,
+	},
+};
+
 static int gic_of_setup(struct gic_chip_data *gic, struct device_node *node)
 {
 	if (!gic || !node)
@@ -1391,6 +1438,8 @@ static int gic_of_setup(struct gic_chip_data *gic, struct device_node *node)
 	if (of_property_read_u32(node, "cpu-offset", &gic->percpu_offset))
 		gic->percpu_offset = 0;
 
+	gic_enable_of_quirks(node, gic_quirks, gic);
+
 	return 0;
 
 error:

-- 
Without deviation from the norm, progress is not possible.