Skip to contentSkip to conversation

Increase the number of IRQ descriptors for SPARSEIRQ

PATCH Superseded Shanker Donthineni 2 participants 7 messages Apr 8, 2023 – Apr 10, 2023 1 message not in the archive IRQ

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v2 cover Increase the number of IRQ descriptors for SPARSEIRQ

The ARM64 architecture uses SPARSEIRQ with a default value of NR_IRQS, which is set to 64. This means that only 64+8192 IRQ descriptors are allowed, which may not be sufficient for modern ARM64 servers that have a large number of IO devices and GIC hardware that supports direct vSGI and vLPI injection features.

This limitation has caused issues when attempting to launch multiple virtual machines with GICv4.1 features, resulting in the error message 'kvm_err("VPE IRQ allocation failure\n")'. The root cause of this issue is the ~8K IRQ descriptor limit.

To address this issue, an initial proposal was made to define NR_IRQS to 2^19 for ARM64. However, Marc Zyngier suggested implementing a generic solution instead of hard-coded values. Thomas Gleixner advised to use the maple tree data structure and provided most of the necessary functions.

For more information, refer to the discussion thread at
https://lore.kernel.org/linux-arm-kernel/[email protected]/.

This patch series converts the static memory allocation to dynamic using the maple tree, and increases the maximum number of IRQ descriptors to INT_MAX from NR_IRQS+8192. This change has been tested on an ARM64 server with CONFIG_SPARSE_IRQ=y, where 256 virtual machines were launched, creating a total of 128K+ IRQ descriptors, and IRQ injection was verified.

The maple_tree RCU mode fixes are prerequisites, expecting all these changes will be available in v6.4-rc1.
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/log/lib/maple_tree.c?h=next-20230406
- maple_tree: add RCU lock checking to rcu callback functions
- maple_tree: add smp_rmb() to dead node detection
- maple_tree: fix write memory barrier of nodes once dead for RCU mode
- maple_tree: remove extra smp_wmb() from mas_dead_leaves()
- maple_tree: fix freeing of nodes in rcu mode
- maple_tree: detect dead nodes in mas_start()
- maple_tree: be more cautious about dead nodes

Changes in v2:
- The patches have been updated to v6.3-rc5.
- The patches 2/5 and 4/5 have been removed as they are unnecessary.
- The review comments from Thomas have been addressed.

Shanker Donthineni (3):
genirq: Use hlist for managing resend handlers
genirq: Encapsulate sparse bitmap handling
genirq: Use the maple tree for IRQ descriptors management

 include/linux/irqdesc.h |  3 ++
 kernel/irq/chip.c       |  1 +
 kernel/irq/internals.h  |  5 +--
 kernel/irq/irqdesc.c    | 80 ++++++++++++++++++++++++++---------------
 kernel/irq/resend.c     | 41 ++++++++++++---------
 5 files changed, 82 insertions(+), 48 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 1/3 genirq: Use hlist for managing resend handlers

The current implementation utilizes a bitmap for managing IRQ resend handlers, which is allocated based on the SPARSE_IRQ/NR_IRQS macros. However, this method may not efficiently utilize memory during runtime, particularly when IRQ_BITMAP_BITS is large.

This proposed patch aims to address this issue by using hlist to manage IRQ resend handlers instead of relying on static memory allocation. Additionally, a new function, clear_irq_resend(), is introduced and called from irq_shutdown to ensure a graceful teardown of IRQD.

---
 include/linux/irqdesc.h |  3 +++
 kernel/irq/chip.c       |  1 +
 kernel/irq/internals.h  |  1 +
 kernel/irq/irqdesc.c    |  6 ++++++
 kernel/irq/resend.c     | 41 ++++++++++++++++++++++++-----------------
 5 files changed, 35 insertions(+), 17 deletions(-)
include/linux/irqdesc.h +3 -0
index 844a8e30e6de..d9451d456a73 100644
@@ -102,6 +102,9 @@ struct irq_desc {
 	int			parent_irq;
 	struct module		*owner;
 	const char		*name;
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+	struct hlist_node	resend_node;
+#endif
 } ____cacheline_internodealigned_in_smp;
 
 #ifdef CONFIG_SPARSE_IRQ
kernel/irq/chip.c +1 -0
index 49e7bc871fec..2eac5532c3c8 100644
@@ -306,6 +306,7 @@ static void __irq_disable(struct irq_desc *desc, bool mask);
 void irq_shutdown(struct irq_desc *desc)
 {
 	if (irqd_is_started(&desc->irq_data)) {
+		clear_irq_resend(desc);
 		desc->depth = 1;
 		if (desc->irq_data.chip->irq_shutdown) {
 			desc->irq_data.chip->irq_shutdown(&desc->irq_data);
kernel/irq/internals.h +1 -0
index 5fdc0b557579..2fd17057ed4b 100644
@@ -113,6 +113,7 @@ irqreturn_t handle_irq_event(struct irq_desc *desc);
 
 /* Resending of interrupts :*/
 int check_irq_resend(struct irq_desc *desc, bool inject);
+void clear_irq_resend(struct irq_desc *desc);
 bool irq_wait_for_poll(struct irq_desc *desc);
 void __irq_wake_thread(struct irq_desc *desc, struct irqaction *action);
 
kernel/irq/irqdesc.c +6 -0
index 240e145e969f..47543b5a0edb 100644
@@ -415,6 +415,9 @@ static struct irq_desc *alloc_desc(int irq, int node, unsigned int flags,
 	desc_set_defaults(irq, desc, node, affinity, owner);
 	irqd_set(&desc->irq_data, flags);
 	kobject_init(&desc->kobj, &irq_kobj_type);
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+	INIT_HLIST_NODE(&desc->resend_node);
+#endif
 
 	return desc;
 
@@ -581,6 +584,9 @@ int __init early_irq_init(void)
 		mutex_init(&desc[i].request_mutex);
 		init_waitqueue_head(&desc[i].wait_for_threads);
 		desc_set_defaults(i, &desc[i], node, NULL, NULL);
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+		INIT_HLIST_NODE(&desc->resend_node);
+#endif
 	}
 	return arch_early_irq_init();
 }
kernel/irq/resend.c +24 -17

74 lines not on this page · raw message

Signature · 1 line
2.25.1

A message that is not in the archive (sent off-list or not yet received): [email protected]

#3 lore raw

Re: [PATCH v2 1/3] genirq: Use hlist for managing resend handlers

Hi Marc,

Quoted 13 lines · Marc Zyngier

External email: Use caution opening links or attachments

Quoted 7 lines · Shanker Donthineni

The current implementation utilizes a bitmap for managing IRQ resend handlers, which is allocated based on the SPARSE_IRQ/NR_IRQS macros. However, this method may not efficiently utilize memory during runtime, particularly when IRQ_BITMAP_BITS is large.

This proposed patch aims to address this issue by using hlist to manage

Quoted 8 lines

s/This proposed patch aims to//

Quoted 3 lines

IRQ resend handlers instead of relying on static memory allocation. Additionally, a new function, clear_irq_resend(), is introduced and called from irq_shutdown to ensure a graceful teardown of IRQD.

s/IRQD/the interrupt/

Ack

Quoted 119 lines
Quoted 53 lines
---
  include/linux/irqdesc.h |  3 +++
  kernel/irq/chip.c       |  1 +
  kernel/irq/internals.h  |  1 +
  kernel/irq/irqdesc.c    |  6 ++++++
  kernel/irq/resend.c     | 41 ++++++++++++++++++++++++-----------------
  5 files changed, 35 insertions(+), 17 deletions(-)
include/linux/irqdesc.h +3 -0
index 844a8e30e6de..d9451d456a73 100644
@@ -102,6 +102,9 @@ struct irq_desc {
       int                     parent_irq;
       struct module           *owner;
       const char              *name;
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+     struct hlist_node       resend_node;
+#endif
  } ____cacheline_internodealigned_in_smp;
 
  #ifdef CONFIG_SPARSE_IRQ
kernel/irq/chip.c +1 -0
index 49e7bc871fec..2eac5532c3c8 100644
@@ -306,6 +306,7 @@ static void __irq_disable(struct irq_desc *desc, bool mask);
  void irq_shutdown(struct irq_desc *desc)
  {
       if (irqd_is_started(&desc->irq_data)) {
+             clear_irq_resend(desc);
               desc->depth = 1;
               if (desc->irq_data.chip->irq_shutdown) {
                       desc->irq_data.chip->irq_shutdown(&desc->irq_data);
kernel/irq/internals.h +1 -0
index 5fdc0b557579..2fd17057ed4b 100644
@@ -113,6 +113,7 @@ irqreturn_t handle_irq_event(struct irq_desc *desc);
 
  /* Resending of interrupts :*/
  int check_irq_resend(struct irq_desc *desc, bool inject);
+void clear_irq_resend(struct irq_desc *desc);
  bool irq_wait_for_poll(struct irq_desc *desc);
  void __irq_wake_thread(struct irq_desc *desc, struct irqaction *action);
 
kernel/irq/irqdesc.c +0 -0
index 240e145e969f..47543b5a0edb 100644

kernel/irq/irqdesc.c static struct irq_desc *alloc_desc(int irq, int node, unsigned int flags, · line 420

Quoted 6 lines
kernel/irq/irqdesc.c +3 -0
@@ -415,6 +415,9 @@ static struct irq_desc *alloc_desc(int irq, int node, unsigned int flags,
       desc_set_defaults(irq, desc, node, affinity, owner);
       irqd_set(&desc->irq_data, flags);
       kobject_init(&desc->kobj, &irq_kobj_type);
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+     INIT_HLIST_NODE(&desc->resend_node);
+#endif

Please add a helper that hides the #ifdefery from the core code, and use it in both spots.
Sure, I will add a helper function irq_resend_init()

Quoted 42 lines
kernel/irq/irqdesc.c +3 -0
       return desc;
 
@@ -581,6 +584,9 @@ int __init early_irq_init(void)
               mutex_init(&desc[i].request_mutex);
               init_waitqueue_head(&desc[i].wait_for_threads);
               desc_set_defaults(i, &desc[i], node, NULL, NULL);
+#ifdef CONFIG_HARDIRQS_SW_RESEND
+             INIT_HLIST_NODE(&desc->resend_node);
+#endif
       }
       return arch_early_irq_init();
  }
kernel/irq/resend.c +3 -11
index 0c46e9fe3a89..d3db2628a720 100644
@@ -21,8 +21,9 @@
 
  #ifdef CONFIG_HARDIRQS_SW_RESEND
 
-/* Bitmap to handle software resend of interrupts: */
-static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS);
+/* hlist_head to handle software resend of interrupts: */
+static HLIST_HEAD(irq_resend_list);
+static DEFINE_RAW_SPINLOCK(irq_resend_lock);
 
  /*
   * Run software resends of IRQ's
@@ -30,18 +31,17 @@ static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS);
  static void resend_irqs(struct tasklet_struct *unused)
  {
       struct irq_desc *desc;
-     int irq;
-
-     while (!bitmap_empty(irqs_resend, nr_irqs)) {
-             irq = find_first_bit(irqs_resend, nr_irqs);
-             clear_bit(irq, irqs_resend);
-             desc = irq_to_desc(irq);
-             if (!desc)
-                     continue;
-             local_irq_disable();

kernel/irq/resend.c static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS); · line 42

Quoted 10 lines
kernel/irq/resend.c +8 -1
+
+     raw_spin_lock_irq(&irq_resend_lock);
+     while (!hlist_empty(&irq_resend_list)) {
+             desc = hlist_entry(irq_resend_list.first, struct irq_desc,
+                                resend_node);
+             hlist_del_init(&desc->resend_node);
+             raw_spin_unlock(&irq_resend_lock);
               desc->handle_irq(desc);
-             local_irq_enable();
+             raw_spin_lock(&irq_resend_lock);

What makes it safe to drop the local_irq_*able()?

Quoted 4 lines

tasklet_action_common() explicitly enables interrupts when calling the callback, so unless there is some other interrupt disabling that I can't immediately spot, the handler may run in the wrong context.

Unless I am overlooking something, interrupts are disabled within the while loop unless desc->handle_irq() is enabling it. The existing code disables and enables interrupts for each handler invocation, whereas the modified code does it only once for all outstanding handlers.

Please let me know if the below code changes are acceptable. Similar to the current interrupt disable/enable logic except the protection lock resend.

static void resend_irqs(struct tasklet_struct *unused)

 {
        struct irq_desc *desc;
-       int irq;
-
-       while (!bitmap_empty(irqs_resend, nr_irqs)) {
-               irq = find_first_bit(irqs_resend, nr_irqs);
-               clear_bit(irq, irqs_resend);
-               desc = irq_to_desc(irq);
-               if (!desc)

- continue;
+
+ raw_spin_lock(&irq_resend_lock);
+ while (!hlist_empty(&irq_resend_list)) {
+ desc = hlist_entry(irq_resend_list.first, struct irq_desc,
+ resend_node);
+ hlist_del_init(&desc->resend_node);
+ raw_spin_unlock(&irq_resend_lock);
+

                local_irq_disable();
                desc->handle_irq(desc);
                local_irq_enable();
+
+               raw_spin_lock(&irq_resend_lock);
        }

+ raw_spin_unlock(&irq_resend_lock);
}

Quoted 7 lines

If there is such interrupt disabling, then please point it out in the commit message so that idiots like me can refer to it.

         M.

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

#4 lore raw

Re: [PATCH v2 1/3] genirq: Use hlist for managing resend handlers

Quoted 32 lines · Shanker Donthineni
Quoted 27 lines
Quoted 14 lines
kernel/irq/resend.c +0 -9
@@ -30,18 +31,17 @@ static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS);
  static void resend_irqs(struct tasklet_struct *unused)
  {
       struct irq_desc *desc;
-     int irq;
-
-     while (!bitmap_empty(irqs_resend, nr_irqs)) {
-             irq = find_first_bit(irqs_resend, nr_irqs);
-             clear_bit(irq, irqs_resend);
-             desc = irq_to_desc(irq);
-             if (!desc)
-                     continue;
-             local_irq_disable();

kernel/irq/resend.c static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS); · line 42

Quoted 10 lines
kernel/irq/resend.c +8 -1
+
+     raw_spin_lock_irq(&irq_resend_lock);
+     while (!hlist_empty(&irq_resend_list)) {
+             desc = hlist_entry(irq_resend_list.first, struct irq_desc,
+                                resend_node);
+             hlist_del_init(&desc->resend_node);
+             raw_spin_unlock(&irq_resend_lock);
               desc->handle_irq(desc);
-             local_irq_enable();
+             raw_spin_lock(&irq_resend_lock);

What makes it safe to drop the local_irq_*able()?

Quoted 4 lines

tasklet_action_common() explicitly enables interrupts when calling the callback, so unless there is some other interrupt disabling that I can't immediately spot, the handler may run in the wrong context.

Quoted 4 lines

Unless I am overlooking something, interrupts are disabled within the while loop unless desc->handle_irq() is enabling it. The existing code disables and enables interrupts for each handler invocation, whereas the modified code does it only once for all outstanding handlers.

Ah, you use raw_spinlock_irq() outside of the loop. I somehow glanced over that, apologies for the noise. Unless we expect a really long list of interrupts to be resent, your current code should be OK.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#5 lore raw

Re: [PATCH v2 1/3] genirq: Use hlist for managing resend handlers

Quoted 37 lines · Marc Zyngier

External email: Use caution opening links or attachments

Quoted 32 lines · Shanker Donthineni
Quoted 31 lines
Quoted 14 lines
kernel/irq/resend.c +0 -9
@@ -30,18 +31,17 @@ static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS);
   static void resend_irqs(struct tasklet_struct *unused)
   {
        struct irq_desc *desc;
-     int irq;
-
-     while (!bitmap_empty(irqs_resend, nr_irqs)) {
-             irq = find_first_bit(irqs_resend, nr_irqs);
-             clear_bit(irq, irqs_resend);
-             desc = irq_to_desc(irq);
-             if (!desc)
-                     continue;
-             local_irq_disable();

kernel/irq/resend.c static DECLARE_BITMAP(irqs_resend, IRQ_BITMAP_BITS); · line 42

Quoted 10 lines
kernel/irq/resend.c +8 -1
+
+     raw_spin_lock_irq(&irq_resend_lock);
+     while (!hlist_empty(&irq_resend_list)) {
+             desc = hlist_entry(irq_resend_list.first, struct irq_desc,
+                                resend_node);
+             hlist_del_init(&desc->resend_node);
+             raw_spin_unlock(&irq_resend_lock);
                desc->handle_irq(desc);
-             local_irq_enable();
+             raw_spin_lock(&irq_resend_lock);

What makes it safe to drop the local_irq_*able()?

tasklet_action_common() explicitly enables interrupts when calling the callback, so unless there is some other interrupt disabling that I can't immediately spot, the handler may run in the wrong context.

Quoted 9 lines
Quoted 4 lines

Unless I am overlooking something, interrupts are disabled within the while loop unless desc->handle_irq() is enabling it. The existing code disables and enables interrupts for each handler invocation, whereas the modified code does it only once for all outstanding handlers.

Ah, you use raw_spinlock_irq() outside of the loop. I somehow glanced over that, apologies for the noise. Unless we expect a really long list of interrupts to be resent, your current code should be OK.

Thanks, I'll post v3 patches to address the other review comments.

#6 lore raw

PATCH v2 2/3 genirq: Encapsulate sparse bitmap handling

In reply to Shanker Donthineni

Move the open coded sparse bitmap handling into helper functions as a preparatory step for converting the sparse interrupt management to a maple tree.

No functional change.

---
 kernel/irq/internals.h |  4 ++--
 kernel/irq/irqdesc.c   | 28 +++++++++++++++++++---------
 2 files changed, 21 insertions(+), 11 deletions(-)
kernel/irq/internals.h +2 -2
index 2fd17057ed4b..5d741b0e7d5e 100644
@@ -12,9 +12,9 @@
 #include <linux/sched/clock.h>
 
 #ifdef CONFIG_SPARSE_IRQ
-# define IRQ_BITMAP_BITS	(NR_IRQS + 8196)
+# define MAX_SPARSE_IRQS	(NR_IRQS + 8196)
 #else
-# define IRQ_BITMAP_BITS	NR_IRQS
+# define MAX_SPARSE_IRQS	NR_IRQS
 #endif
 
 #define istate core_internal_state__do_not_mess_with_it
kernel/irq/irqdesc.c +19 -9
index 47543b5a0edb..e0c259769d3a 100644
@@ -131,7 +131,18 @@ int nr_irqs = NR_IRQS;
 EXPORT_SYMBOL_GPL(nr_irqs);
 
 static DEFINE_MUTEX(sparse_irq_lock);
-static DECLARE_BITMAP(allocated_irqs, IRQ_BITMAP_BITS);
+static DECLARE_BITMAP(allocated_irqs, MAX_SPARSE_IRQS);
+
+static int irq_find_free_area(unsigned int from, unsigned int cnt)
+{
+	return bitmap_find_next_zero_area(allocated_irqs, MAX_SPARSE_IRQS,
+					  from, cnt, 0);
+}
+
+static unsigned int irq_find_next_irq(unsigned int offset)
+{
+	return find_next_bit(allocated_irqs, nr_irqs, offset);
+}
 
 #ifdef CONFIG_SPARSE_IRQ
 
@@ -519,7 +530,7 @@ static int alloc_descs(unsigned int start, unsigned int cnt, int node,
 
 static int irq_expand_nr_irqs(unsigned int nr)
 {
-	if (nr > IRQ_BITMAP_BITS)
+	if (nr > MAX_SPARSE_IRQS)
 		return -ENOMEM;
 	nr_irqs = nr;
 	return 0;
@@ -537,11 +548,11 @@ int __init early_irq_init(void)
 	printk(KERN_INFO "NR_IRQS: %d, nr_irqs: %d, preallocated irqs: %d\n",
 	       NR_IRQS, nr_irqs, initcnt);
 
-	if (WARN_ON(nr_irqs > IRQ_BITMAP_BITS))
-		nr_irqs = IRQ_BITMAP_BITS;
+	if (WARN_ON(nr_irqs > MAX_SPARSE_IRQS))
+		nr_irqs = MAX_SPARSE_IRQS;
 
-	if (WARN_ON(initcnt > IRQ_BITMAP_BITS))
-		initcnt = IRQ_BITMAP_BITS;
+	if (WARN_ON(initcnt > MAX_SPARSE_IRQS))
+		initcnt = MAX_SPARSE_IRQS;
 
 	if (initcnt > nr_irqs)
 		nr_irqs = initcnt;
@@ -816,8 +827,7 @@ __irq_alloc_descs(int irq, unsigned int from, unsigned int cnt, int node,
 
 	mutex_lock(&sparse_irq_lock);
 
-	start = bitmap_find_next_zero_area(allocated_irqs, IRQ_BITMAP_BITS,
-					   from, cnt, 0);
+	start = irq_find_free_area(from, cnt);
 	ret = -EEXIST;
 	if (irq >=0 && start != irq)
 		goto unlock;
@@ -842,7 +852,7 @@ EXPORT_SYMBOL_GPL(__irq_alloc_descs);
  */
 unsigned int irq_get_next_irq(unsigned int offset)
 {
-	return find_next_bit(allocated_irqs, nr_irqs, offset);
+	return irq_find_next_irq(offset);
 }
 
 struct irq_desc *
Signature · 1 line
2.25.1
#7 lore raw

PATCH v2 3/3 genirq: Use the maple tree for IRQ descriptors management

In reply to Shanker Donthineni

The current implementation uses a static bitmap and a radix tree to manage IRQ allocation and irq_desc pointer store respectively. However, the size of the bitmap is constrained by the build time macro MAX_SPARSE_IRQS, which may not be sufficient to support the high-end servers, particularly those with GICv4.1 hardware, which require a large interrupt space to cover LPIs and vSGIs

The maple tree is a highly efficient data structure for storing non-overlapping ranges and can handle a large number of entries, up to ULONG_MAX. It can be utilized for both storing IRQD and identifying available free spaces.

The interrupt descriptors management can be simplified by switching to a maple tree data structure, which offers greater flexibility and scalability. To support modern servers, the maximum number of IRQs has been increased to INT_MAX, which provides a more adequate value than the previous limit of NR_IRQS+8192.

---
 kernel/irq/internals.h |  2 +-
 kernel/irq/irqdesc.c   | 54 +++++++++++++++++++++++-------------------
 2 files changed, 31 insertions(+), 25 deletions(-)
kernel/irq/internals.h +1 -1
index 5d741b0e7d5e..e35de737802c 100644
@@ -12,7 +12,7 @@
 #include <linux/sched/clock.h>
 
 #ifdef CONFIG_SPARSE_IRQ
-# define MAX_SPARSE_IRQS	(NR_IRQS + 8196)
+# define MAX_SPARSE_IRQS	INT_MAX
 #else
 # define MAX_SPARSE_IRQS	NR_IRQS
 #endif
kernel/irq/irqdesc.c +30 -24

123 lines not on this page · raw message

Signature · 1 line
2.25.1