[PATCH 0/2] irqchip/gic-v3-its: a raw_spinlock_t + early ->pend_page allocateion

[PATCH 0/2] irqchip/gic-v3-its: a raw_spinlock_t + early ->pend_page allocateion

[Sebastian Andrzej Siewior]

The errors popped up in -RT. One is the spinlock_t usage in IRQ-off
region which should be a raw_spinlock_t. I don't see this to be used
during run-time or for long period of time so I don't thing this is an
issue.

The second problem is the ->pend_page allocation on the target_cpu early
during CPU bring up. It would be simpler to allocate the memory once on
the boot CPU instead the hotplug callback since the memory is never
free()ed. I *think* this is avoided because the number of possible CPUs
might be higher than what will actually come up (on x86 the BIOS
sometimes reports more possible_cpus than what is *really* possible, not
sure if ARM64 is there yet). Therefore I moved the allocation to an
earlier callback.

Sebastian

[PATCH 1/2] irqchip/gic-v3-its: Make its_lock a raw_spin_lock_t

[Sebastian Andrzej Siewior]

The its_lock lock is held while a new device is added to the list and
during setup while the CPU is booted. Even on -RT the CPU-bootup is
performed with disabled interrupts.

Make its_lock a raw_spin_lock_t.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 drivers/irqchip/irq-gic-v3-its.c | 18 +++++++++---------
 1 file changed, 9 insertions(+), 9 deletions(-)

diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index d7842d312d3e..a616043d25ee 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -160,7 +160,7 @@ static struct {
 } vpe_proxy;
 
 static LIST_HEAD(its_nodes);
-static DEFINE_SPINLOCK(its_lock);
+static DEFINE_RAW_SPINLOCK(its_lock);
 static struct rdists *gic_rdists;
 static struct irq_domain *its_parent;
 
@@ -1997,12 +1997,12 @@ static void its_cpu_init_collections(void)
 {
 	struct its_node *its;
 
-	spin_lock(&its_lock);
+	raw_spin_lock(&its_lock);
 
 	list_for_each_entry(its, &its_nodes, entry)
 		its_cpu_init_collection(its);
 
-	spin_unlock(&its_lock);
+	raw_spin_unlock(&its_lock);
 }
 
 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
@@ -3070,7 +3070,7 @@ static int its_save_disable(void)
 	struct its_node *its;
 	int err = 0;
 
-	spin_lock(&its_lock);
+	raw_spin_lock(&its_lock);
 	list_for_each_entry(its, &its_nodes, entry) {
 		void __iomem *base;
 
@@ -3102,7 +3102,7 @@ static int its_save_disable(void)
 			writel_relaxed(its->ctlr_save, base + GITS_CTLR);
 		}
 	}
-	spin_unlock(&its_lock);
+	raw_spin_unlock(&its_lock);
 
 	return err;
 }
@@ -3112,7 +3112,7 @@ static void its_restore_enable(void)
 	struct its_node *its;
 	int ret;
 
-	spin_lock(&its_lock);
+	raw_spin_lock(&its_lock);
 	list_for_each_entry(its, &its_nodes, entry) {
 		void __iomem *base;
 		int i;
@@ -3164,7 +3164,7 @@ static void its_restore_enable(void)
 		    GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
 			its_cpu_init_collection(its);
 	}
-	spin_unlock(&its_lock);
+	raw_spin_unlock(&its_lock);
 }
 
 static struct syscore_ops its_syscore_ops = {
@@ -3398,9 +3398,9 @@ static int __init its_probe_one(struct resource *res,
 	if (err)
 		goto out_free_tables;
 
-	spin_lock(&its_lock);
+	raw_spin_lock(&its_lock);
 	list_add(&its->entry, &its_nodes);
-	spin_unlock(&its_lock);
+	raw_spin_unlock(&its_lock);
 
 	return 0;
 
-- 
2.18.0

[PATCH 2/2] irqchip/gic-v3-its: Move ITS' ->pend_page allocation into an early CPU up callback

[Sebastian Andrzej Siewior]

The AP-GIC-starting callback allocates memory for the ->pend_page while
the CPU is started during boot-up. This callback is invoked on the
target CPU with disabled interrupts.
This does not work on -RT because memory allocations are not possible
with disabled interrupts.
Move the memory allocation to an earlier hotplug step which is invoked
with enabled interrupts on the boot CPU.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
---
 drivers/irqchip/irq-gic-v3-its.c | 60 ++++++++++++++++++++++----------
 1 file changed, 41 insertions(+), 19 deletions(-)

diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index a616043d25ee..acc3d44c356d 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -171,6 +171,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock);
 static DEFINE_IDA(its_vpeid_ida);
 
 #define gic_data_rdist()		(raw_cpu_ptr(gic_rdists->rdist))
+#define gic_data_rdist_cpu(cpu)		(per_cpu_ptr(gic_rdists->rdist, cpu))
 #define gic_data_rdist_rd_base()	(gic_data_rdist()->rd_base)
 #define gic_data_rdist_vlpi_base()	(gic_data_rdist_rd_base() + SZ_128K)
 
@@ -1853,15 +1854,17 @@ static int its_alloc_collections(struct its_node *its)
 	return 0;
 }
 
-static struct page *its_allocate_pending_table(gfp_t gfp_flags)
+static struct page *its_allocate_pending_table(unsigned int cpu)
 {
 	struct page *pend_page;
+	unsigned int order;
 	/*
 	 * The pending pages have to be at least 64kB aligned,
 	 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
 	 */
-	pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
-				get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
+	order = get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K));
+	pend_page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_ZERO,
+				     order);
 	if (!pend_page)
 		return NULL;
 
@@ -1877,6 +1880,28 @@ static void its_free_pending_table(struct page *pt)
 		   get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
 }
 
+static int its_alloc_pend_page(unsigned int cpu)
+{
+	struct page *pend_page;
+	phys_addr_t paddr;
+
+	pend_page = gic_data_rdist_cpu(cpu)->pend_page;
+	if (pend_page)
+		return 0;
+
+	pend_page = its_allocate_pending_table(cpu);
+	if (!pend_page) {
+		pr_err("Failed to allocate PENDBASE for CPU%d\n",
+		       smp_processor_id());
+		return -ENOMEM;
+	}
+
+	paddr = page_to_phys(pend_page);
+	pr_info("CPU%d: using LPI pending table @%pa\n", cpu, &paddr);
+	gic_data_rdist_cpu(cpu)->pend_page = pend_page;
+	return 0;
+}
+
 static void its_cpu_init_lpis(void)
 {
 	void __iomem *rbase = gic_data_rdist_rd_base();
@@ -1885,21 +1910,8 @@ static void its_cpu_init_lpis(void)
 
 	/* If we didn't allocate the pending table yet, do it now */
 	pend_page = gic_data_rdist()->pend_page;
-	if (!pend_page) {
-		phys_addr_t paddr;
-
-		pend_page = its_allocate_pending_table(GFP_NOWAIT);
-		if (!pend_page) {
-			pr_err("Failed to allocate PENDBASE for CPU%d\n",
-			       smp_processor_id());
-			return;
-		}
-
-		paddr = page_to_phys(pend_page);
-		pr_info("CPU%d: using LPI pending table @%pa\n",
-			smp_processor_id(), &paddr);
-		gic_data_rdist()->pend_page = pend_page;
-	}
+	if (WARN_ON(!pend_page))
+		return;
 
 	/* set PROPBASE */
 	val = (page_to_phys(gic_rdists->prop_page) |
@@ -2732,7 +2744,7 @@ static int its_vpe_init(struct its_vpe *vpe)
 		return vpe_id;
 
 	/* Allocate VPT */
-	vpt_page = its_allocate_pending_table(GFP_KERNEL);
+	vpt_page = its_allocate_pending_table(raw_smp_processor_id());
 	if (!vpt_page) {
 		its_vpe_id_free(vpe_id);
 		return -ENOMEM;
@@ -3706,6 +3718,16 @@ int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
 	if (err)
 		return err;
 
+	err = cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "irqchip/arm/gicv3:prepare",
+				its_alloc_pend_page, NULL);
+	if (err < 0) {
+		pr_warn("ITS: Can't register CPU-hoplug callback.\n");
+		return err;
+	}
+	err = its_alloc_pend_page(smp_processor_id());
+	if (err < 0)
+		return err;
+
 	list_for_each_entry(its, &its_nodes, entry)
 		has_v4 |= its->is_v4;
 
-- 
2.18.0

Re: [PATCH 2/2] irqchip/gic-v3-its: Move ITS' ->pend_page allocation into an early CPU up callback

[Marc Zyngier]

Hi Sebastian,

On 18/07/18 16:42, Sebastian Andrzej Siewior wrote:
> The AP-GIC-starting callback allocates memory for the ->pend_page while
> the CPU is started during boot-up. This callback is invoked on the
> target CPU with disabled interrupts.
> This does not work on -RT because memory allocations are not possible
> with disabled interrupts.
> Move the memory allocation to an earlier hotplug step which is invoked
> with enabled interrupts on the boot CPU.
> 
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
>  drivers/irqchip/irq-gic-v3-its.c | 60 ++++++++++++++++++++++----------
>  1 file changed, 41 insertions(+), 19 deletions(-)
> 
> diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
> index a616043d25ee..acc3d44c356d 100644
> --- a/drivers/irqchip/irq-gic-v3-its.c
> +++ b/drivers/irqchip/irq-gic-v3-its.c
> @@ -171,6 +171,7 @@ static DEFINE_RAW_SPINLOCK(vmovp_lock);
>  static DEFINE_IDA(its_vpeid_ida);
>  
>  #define gic_data_rdist()		(raw_cpu_ptr(gic_rdists->rdist))
> +#define gic_data_rdist_cpu(cpu)		(per_cpu_ptr(gic_rdists->rdist, cpu))
>  #define gic_data_rdist_rd_base()	(gic_data_rdist()->rd_base)
>  #define gic_data_rdist_vlpi_base()	(gic_data_rdist_rd_base() + SZ_128K)
>  
> @@ -1853,15 +1854,17 @@ static int its_alloc_collections(struct its_node *its)
>  	return 0;
>  }
>  
> -static struct page *its_allocate_pending_table(gfp_t gfp_flags)
> +static struct page *its_allocate_pending_table(unsigned int cpu)
>  {
>  	struct page *pend_page;
> +	unsigned int order;
>  	/*
>  	 * The pending pages have to be at least 64kB aligned,
>  	 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
>  	 */
> -	pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
> -				get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
> +	order = get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K));
> +	pend_page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL | __GFP_ZERO,
> +				     order);
>  	if (!pend_page)
>  		return NULL;
>  
> @@ -1877,6 +1880,28 @@ static void its_free_pending_table(struct page *pt)
>  		   get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
>  }
>  
> +static int its_alloc_pend_page(unsigned int cpu)
> +{
> +	struct page *pend_page;
> +	phys_addr_t paddr;
> +
> +	pend_page = gic_data_rdist_cpu(cpu)->pend_page;
> +	if (pend_page)
> +		return 0;
> +
> +	pend_page = its_allocate_pending_table(cpu);
> +	if (!pend_page) {
> +		pr_err("Failed to allocate PENDBASE for CPU%d\n",
> +		       smp_processor_id());
> +		return -ENOMEM;
> +	}
> +
> +	paddr = page_to_phys(pend_page);
> +	pr_info("CPU%d: using LPI pending table @%pa\n", cpu, &paddr);
> +	gic_data_rdist_cpu(cpu)->pend_page = pend_page;
> +	return 0;
> +}
> +
>  static void its_cpu_init_lpis(void)
>  {
>  	void __iomem *rbase = gic_data_rdist_rd_base();
> @@ -1885,21 +1910,8 @@ static void its_cpu_init_lpis(void)
>  
>  	/* If we didn't allocate the pending table yet, do it now */
>  	pend_page = gic_data_rdist()->pend_page;
> -	if (!pend_page) {
> -		phys_addr_t paddr;
> -
> -		pend_page = its_allocate_pending_table(GFP_NOWAIT);
> -		if (!pend_page) {
> -			pr_err("Failed to allocate PENDBASE for CPU%d\n",
> -			       smp_processor_id());
> -			return;
> -		}
> -
> -		paddr = page_to_phys(pend_page);
> -		pr_info("CPU%d: using LPI pending table @%pa\n",
> -			smp_processor_id(), &paddr);
> -		gic_data_rdist()->pend_page = pend_page;
> -	}
> +	if (WARN_ON(!pend_page))
> +		return;
>  
>  	/* set PROPBASE */
>  	val = (page_to_phys(gic_rdists->prop_page) |
> @@ -2732,7 +2744,7 @@ static int its_vpe_init(struct its_vpe *vpe)
>  		return vpe_id;
>  
>  	/* Allocate VPT */
> -	vpt_page = its_allocate_pending_table(GFP_KERNEL);
> +	vpt_page = its_allocate_pending_table(raw_smp_processor_id());
>  	if (!vpt_page) {
>  		its_vpe_id_free(vpe_id);
>  		return -ENOMEM;
> @@ -3706,6 +3718,16 @@ int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
>  	if (err)
>  		return err;
>  
> +	err = cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "irqchip/arm/gicv3:prepare",
> +				its_alloc_pend_page, NULL);
> +	if (err < 0) {
> +		pr_warn("ITS: Can't register CPU-hoplug callback.\n");
> +		return err;
> +	}
> +	err = its_alloc_pend_page(smp_processor_id());
> +	if (err < 0)
> +		return err;
> +
>  	list_for_each_entry(its, &its_nodes, entry)
>  		has_v4 |= its->is_v4;
>  
> 

I must say I'm not overly keen on the extra hotplug notifier, and I'd
rather allocate all the memory upfront on the boot CPU.

I've implemented this as part of a series that is required for
kexec/kdump support[1] (and specifically this[2] patch). Could you
please have a look and let me know if that works for you? I otherwise
plan to post it for review once the 4.19 merge window has closed.

Thanks,

	M.

[1]
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=irq/gicv3-kdump
[2]
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/commit/?h=irq/gicv3-kdump&id=effe377d415900af10e206173d60dc6cc3f0662c
-- 
Jazz is not dead. It just smells funny...

Re: [PATCH 1/2] irqchip/gic-v3-its: Make its_lock a raw_spin_lock_t

[Marc Zyngier]

On 18/07/18 16:42, Sebastian Andrzej Siewior wrote:
> The its_lock lock is held while a new device is added to the list and
> during setup while the CPU is booted. Even on -RT the CPU-bootup is
> performed with disabled interrupts.
> 
> Make its_lock a raw_spin_lock_t.
> 
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

Applied to irq/irqchip-next.

Thanks,
	
	M.
-- 
Jazz is not dead. It just smells funny...

Re: [PATCH 2/2] irqchip/gic-v3-its: Move ITS' ->pend_page allocation into an early CPU up callback

[Sebastian Andrzej Siewior]

On 2018-08-06 09:15:32 [+0100], Marc Zyngier wrote:
> Hi Sebastian,
Hi Marc,

> I must say I'm not overly keen on the extra hotplug notifier, and I'd
> rather allocate all the memory upfront on the boot CPU.
> 
> I've implemented this as part of a series that is required for
> kexec/kdump support[1] (and specifically this[2] patch). Could you
> please have a look and let me know if that works for you? I otherwise
> plan to post it for review once the 4.19 merge window has closed.

This looks good, thank you. I can't test this because I don't have a
GICv3 box atm. The last one I had died somehow.
I pull that patch into my next RT release and ask for some feedback.
One question: You have
	its_allocate_pending_table(GFP_NOWAIT);

for the allocation. Shouldn't GFP_KERNEL do the job, too? 

> Thanks,
> 
> 	M.

Sebastian

Re: [PATCH 2/2] irqchip/gic-v3-its: Move ITS' ->pend_page allocation into an early CPU up callback

[Sebastian Andrzej Siewior]

On 2018-08-29 18:49:06 [+0200], To Marc Zyngier wrote:
> On 2018-08-06 09:15:32 [+0100], Marc Zyngier wrote:
> > Hi Sebastian,
Hi Marc,

> > I've implemented this as part of a series that is required for
> > kexec/kdump support[1] (and specifically this[2] patch). Could you
> > please have a look and let me know if that works for you? I otherwise
> > plan to post it for review once the 4.19 merge window has closed.

> I pull that patch into my next RT release and ask for some feedback.

Grygorii Strashko confirmed that it works on RT as expected :)

Sebastian