openwrt/target/linux/realtek/patches-5.10/319-irqchip-irq-realtek-rtl-fix-VPE-affinity.patch

160 lines
5.3 KiB
Diff
Raw Normal View History

From 2cd00b51470a30198b048a5fca48a04db77e29cc Mon Sep 17 00:00:00 2001
From: INAGAKI Hiroshi <musashino.open@gmail.com>
Date: Fri, 21 May 2021 23:16:37 +0900
Subject: [PATCH] realtek: backport irq-realtek-rtl driver from 5.12 to 5.10
This patch backports "irq-realtek-rtl" driver to Kernel 5.10 from 5.12.
"MACH_REALTEK_RTL" is used as a platform name in upstream, but "RTL838X"
is used in OpenWrt, so update the dependency by the additional patch.
Submitted-by: INAGAKI Hiroshi <musashino.open@gmail.com>
---
drivers/irqchip/irq-realtek-rtl.c | 38 +++++++++++------
1 files changed, 58 insertions(+), 20 deletions(-)
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
--- a/drivers/irqchip/irq-realtek-rtl.c
+++ b/drivers/irqchip/irq-realtek-rtl.c
@@ -28,6 +28,7 @@ static DEFINE_RAW_SPINLOCK(irq_lock);
#define REG(offset, cpu) (realtek_ictl_base[cpu] + offset)
+static u32 realtek_ictl_unmask[NR_CPUS];
static void __iomem *realtek_ictl_base[NR_CPUS];
static cpumask_t realtek_ictl_cpu_configurable;
@@ -41,11 +42,29 @@ struct realtek_ictl_output {
};
/*
- * IRR0-IRR3 store 4 bits per interrupt, but Realtek uses inverted numbering,
- * placing IRQ 31 in the first four bits. A routing value of '0' means the
- * interrupt is left disconnected. Routing values {1..15} connect to output
- * lines {0..14}.
+ * Per CPU we have a set of 5 registers that determine interrupt handling for
+ * 32 external interrupts. GIMR (enable/disable interrupt) plus IRR0-IRR3 that
+ * contain "routing" or "priority" values. GIMR uses one bit for each interrupt
+ * and IRRx store 4 bits per interrupt. Realtek uses inverted numbering,
+ * placing IRQ 31 in the first four bits. The register combinations give the
+ * following results for a single interrupt in the wild:
+ *
+ * a) GIMR = 0 / IRRx > 0 -> no interrupts
+ * b) GIMR = 0 / IRRx = 0 -> no interrupts
+ * c) GIMR = 1 / IRRx > 0 -> interrupts
+ * d) GIMR = 1 / IRRx = 0 -> rare interrupts in SMP environment
+ *
+ * Combination d) seems to trigger interrupts only on a VPE if the other VPE
+ * has GIMR = 0 and IRRx > 0. E.g. busy without interrupts allowed. To provide
+ * IRQ balancing features in SMP this driver will handle the registers as
+ * follows:
+ *
+ * 1) set IRRx > 0 for VPE where the interrupt is desired
+ * 2) set IRRx = 0 for VPE where the interrupt is not desired
+ * 3) set both GIMR = 0 to mask (disabled) interrupt
+ * 4) set GIMR = 1 to unmask (enable) interrupt but only for VPE where IRRx > 0
*/
+
#define IRR_OFFSET(idx) (4 * (3 - (idx * 4) / 32))
#define IRR_SHIFT(idx) ((idx * 4) % 32)
@@ -65,19 +84,33 @@ static inline void write_irr(void __iome
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
writel(irr, irr0 + offset);
}
+static inline void enable_gimr(int hwirq, int cpu)
+{
+ u32 value;
+
+ value = readl(REG(RTL_ICTL_GIMR, cpu));
+ value |= (BIT(hwirq) & realtek_ictl_unmask[cpu]);
+ writel(value, REG(RTL_ICTL_GIMR, cpu));
+}
+
+static inline void disable_gimr(int hwirq, int cpu)
+{
+ u32 value;
+
+ value = readl(REG(RTL_ICTL_GIMR, cpu));
+ value &= ~BIT(hwirq);
+ writel(value, REG(RTL_ICTL_GIMR, cpu));
+}
+
static void realtek_ictl_unmask_irq(struct irq_data *i)
{
unsigned long flags;
- u32 value;
int cpu;
raw_spin_lock_irqsave(&irq_lock, flags);
- for_each_cpu(cpu, &realtek_ictl_cpu_configurable) {
- value = readl(REG(RTL_ICTL_GIMR, cpu));
- value |= BIT(i->hwirq);
- writel(value, REG(RTL_ICTL_GIMR, cpu));
- }
+ for_each_cpu(cpu, &realtek_ictl_cpu_configurable)
+ enable_gimr(i->hwirq, cpu);
raw_spin_unlock_irqrestore(&irq_lock, flags);
}
@@ -85,16 +118,12 @@ static void realtek_ictl_unmask_irq(stru
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
static void realtek_ictl_mask_irq(struct irq_data *i)
{
unsigned long flags;
- u32 value;
int cpu;
raw_spin_lock_irqsave(&irq_lock, flags);
- for_each_cpu(cpu, &realtek_ictl_cpu_configurable) {
- value = readl(REG(RTL_ICTL_GIMR, cpu));
- value &= ~BIT(i->hwirq);
- writel(value, REG(RTL_ICTL_GIMR, cpu));
- }
+ for_each_cpu(cpu, &realtek_ictl_cpu_configurable)
+ disable_gimr(i->hwirq, cpu);
raw_spin_unlock_irqrestore(&irq_lock, flags);
}
@@ -116,11 +145,17 @@ static int __maybe_unused realtek_ictl_i
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
cpumask_and(&cpu_enable, &cpu_configure, dest);
cpumask_andnot(&cpu_disable, &cpu_configure, dest);
- for_each_cpu(cpu, &cpu_disable)
+ for_each_cpu(cpu, &cpu_disable) {
write_irr(REG(RTL_ICTL_IRR0, cpu), i->hwirq, 0);
+ realtek_ictl_unmask[cpu] &= ~BIT(i->hwirq);
+ disable_gimr(i->hwirq, cpu);
+ }
- for_each_cpu(cpu, &cpu_enable)
+ for_each_cpu(cpu, &cpu_enable) {
write_irr(REG(RTL_ICTL_IRR0, cpu), i->hwirq, output->output_index + 1);
+ realtek_ictl_unmask[cpu] |= BIT(i->hwirq);
+ enable_gimr(i->hwirq, cpu);
+ }
irq_data_update_effective_affinity(i, &cpu_enable);
@@ -149,6 +184,7 @@ static int intc_map(struct irq_domain *d
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
output->child_mask |= BIT(hw);
write_irr(REG(RTL_ICTL_IRR0, 0), hw, output->output_index + 1);
+ realtek_ictl_unmask[0] |= BIT(hw);
raw_spin_unlock_irqrestore(&irq_lock, flags);
@@ -279,9 +315,11 @@ static int __init realtek_rtl_of_init(st
realtek: avoid wrong interrupt routing The interrupt controller depends on two control registers. GIMR enables or disables interrupts and IRRx routes these to MIPS CPU interrupts 2-7. Wiki currently states "A value of '0' (in IRRx) disconnects this input from the output line, independent of the line's setting in GIMR." Contrary to normal intuition this statement DOES NOT mean, that interrupts can be disabled by IRRx alone. The sad truth was discovered by enabling SMP for an Zyxel XGS1010 on the 930x target. It shows that driver and interrupts behave as follows: - Timer 0 interrupt 7 has active routing to CPU0 and no routing to CPU1 - Timer 1 interrupt 8 has no routing to CPU0 and active routing to CPU1 - Unmasking (enabling) interrupts writes 1 bits to all GIMR registers - Masking (disabling) interrupts writes 0 bits to both GIMR registers During operation we can encounter a situation like - GIMR bit for a interrupt/CPU combination is set to enabed (=1) - IRRx routing bits for a interrupt/CPU combination are set to disabed (=0) This setting already allows the hardware to fire interrupts to the target CPU/VPE if the other CPU/VPE is currently busy. Especially for CPU bound timer interrupts this is lethal. If timer interrupt 7 arrives at CPU1 and vice versa for interrupt 8 the restart trigger gets lost. The timer dies and a msleep() operation in the kernel will halt endlessly. Fix this by tracking the IRRx active routing setting in a new bitfield with 0="routing active" and 1="no routing". Enable interrupts in GIMR only for a interrupt & CPU if routing is active. Thus we have - GIMR = 0 / IRRx = 0 -> everything disabled - GIMR = 1 / IRRx > 0 -> active and normal routing - GIMR = 0 / IRRx > 0 -> masked (disabled) with normal routing - GIMR = 1 / IRRx = 0 -> no longer possible Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
2022-09-07 15:02:35 +00:00
cpumask_set_cpu(cpu, &realtek_ictl_cpu_configurable);
/* Disable all cascaded interrupts and clear routing */
- writel(0, REG(RTL_ICTL_GIMR, cpu));
- for (soc_irq = 0; soc_irq < RTL_ICTL_NUM_INPUTS; soc_irq++)
+ for (soc_irq = 0; soc_irq < RTL_ICTL_NUM_INPUTS; soc_irq++) {
write_irr(REG(RTL_ICTL_IRR0, cpu), soc_irq, 0);
+ realtek_ictl_unmask[cpu] &= ~BIT(soc_irq);
+ disable_gimr(soc_irq, cpu);
+ }
}
}