On 7/10/26 10:39, Liao, Chang wrote:

在 2026/7/9 20:13, Vladimir Murzin 写道:

quoted

From: Ada Couprie Diaz <redacted>

The entry code handles interrupt masking differently from the rest of
the kernel. Exception handlers enter and exit with all exceptions
masked, but they must temporarily unmask the appropriate set of
exceptions so that the rest of the handler executes with the expected
exception state.

For EL0 handlers, this means dropping to masking context appropriate
for the work to be performed. For EL1 handlers, this means restoring
the masking context of the interrupted task. In both cases, all
exceptions must be masked again before returning from the exception
handler.

The rest of the kernel typically follows the opposite pattern: it
raises the masking context to protect a critical section and later
restores the previous context.

Given these different usage patterns, introduce a dedicated set of
exception masking helpers for the entry code. Keeping these helpers
separate from the generic interrupt masking APIs makes the intended
usage explicit and helps avoid mixing the two masking models.

To make the masking logic easier to reason about, introduce exception
contexts that map directly to the corresponding hardware exception
state. Along with these contexts, provide helpers to:

- translate an exception context into the corresponding hardware
  state,

- verify that the current hardware exception state matches the
  expected exception context,

- raise or lower the current exception context, and

- perform the common mask/unmask operations when the starting or
  target exception context is already known.

Tracking the current exception context also provides two additional
benefits:

- improved debugging by verifying that the hardware exception state
  matches the expected exception context, and

- avoiding unnecessary writes to the hardware exception state.

Signed-off-by: Ada Couprie Diaz <redacted>
Signed-off-by: Vladimir Murzin <redacted>
---
 .../include/asm/interrupts/common_flags.h     | 197 ++++++++++++++++++
 arch/arm64/include/asm/interrupts/entry.h     | 110 ++++++++++
 2 files changed, 307 insertions(+)
 create mode 100644 arch/arm64/include/asm/interrupts/common_flags.h
 create mode 100644 arch/arm64/include/asm/interrupts/entry.h

diff --git a/arch/arm64/include/asm/interrupts/common_flags.h b/arch/arm64/include/asm/interrupts/common_flags.h
new file mode 100644
index 000000000000..6ce60d1519e8
--- /dev/null
+++ b/arch/arm64/include/asm/interrupts/common_flags.h

@@ -0,0 +1,197 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2025 Arm Ltd.
+ */
+#ifndef __ASM_INTERRUPTS_COMMON_FLAGS_H
+#define __ASM_INTERRUPTS_COMMON_FLAGS_H
+
+#include <asm/arch_gicv3.h>
+#include <asm/bug.h>
+#include <asm/cpufeature.h>
+#include <asm/ptrace.h>
+#include <asm/sysreg.h>
+#include <asm/irqflags.h>
+
+#define DAIF_PROCCTX		0
+#define DAIF_PROCCTX_NOIRQ	(PSR_I_BIT | PSR_F_BIT)
+#define DAIF_ERRCTX		(PSR_A_BIT | PSR_I_BIT | PSR_F_BIT)
+#define DAIF_MASK		(PSR_D_BIT | PSR_A_BIT | PSR_I_BIT | PSR_F_BIT)
+
+/*
+ * Exception context mapping
+ *
+ * pseudo-NMI
+ *
+ * CRITICAL -> DAIF + IRQON  (corresponds to the state on exception entry)
+ * ERROR    ->  AIF + IRQON
+ * NONMI    ->   IF + IRQON
+ * NOIRQ    ->    0 + IRQOFF
+ * PROCESS  ->    0 + IRQON
+ *
+ * Otherwise
+ *
+ * CRITICAL -> DAIF           (corresponds to the state on exception entry)
+ * ERROR    ->  AIF
+ * NONMI    ->   IF
+ * NOIRQ    ->   IF
+ * PROCESS  ->    0
+ */
+typedef enum arm64_exc_context {
+	PROCESS_CONTEXT,
+	NOIRQ_CONTEXT,
+	NONMI_CONTEXT,
+	ERROR_CONTEXT,
+	CRITICAL_CONTEXT,
+} arm64_exc_context_t;
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_exc_hwstate_of_process_context(void)
+{
+	if (system_uses_irq_prio_masking())
+		return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX, .pmr=GIC_PRIO_IRQON};
+
+	return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX};
+}
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_exc_hwstate_of_noirq_context(void)
+{
+	if (system_uses_irq_prio_masking())
+		return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX, .pmr=GIC_PRIO_IRQOFF};
+
+	return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX_NOIRQ};
+}
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_exc_hwstate_of_nonmi_context(void)
+{
+	if (system_uses_irq_prio_masking())
+		return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX_NOIRQ, .pmr=GIC_PRIO_IRQON};
+
+	return (arm64_exc_hwstate_t){.daif=DAIF_PROCCTX_NOIRQ};
+}
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_exc_hwstate_of_error_context(void)
+{
+	if (system_uses_irq_prio_masking())
+		return (arm64_exc_hwstate_t){.daif=DAIF_ERRCTX, .pmr=GIC_PRIO_IRQON};
+
+	return (arm64_exc_hwstate_t){.daif=DAIF_ERRCTX};
+}
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_exc_hwstate_of_critical_context(void)
+{
+	if (system_uses_irq_prio_masking())
+		return (arm64_exc_hwstate_t){.daif=DAIF_MASK, .pmr=GIC_PRIO_IRQON};
+
+	return (arm64_exc_hwstate_t){.daif=DAIF_MASK};
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_exc_hwstate_of_context(arm64_exc_context_t context) {
+	switch (context) {
+	case PROCESS_CONTEXT:
+		return __arm64_exc_hwstate_of_process_context();
+	case NOIRQ_CONTEXT:
+		return __arm64_exc_hwstate_of_noirq_context();
+	case NONMI_CONTEXT:
+		return __arm64_exc_hwstate_of_nonmi_context();
+	case ERROR_CONTEXT:
+		return __arm64_exc_hwstate_of_error_context();
+	case CRITICAL_CONTEXT:
+		return __arm64_exc_hwstate_of_critical_context();
+	default:
+		BUG();
+	}
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_inherit_exc_hwstate(struct pt_regs *regs)
+{
+	arm64_exc_hwstate_t state = {.daif=regs->pstate & DAIF_MASK};
+
+	if (system_uses_irq_prio_masking())
+		state.pmr = regs->pmr;
+
+	return state;
+}
+
+static __always_inline
+void arm64_debug_exc_hwstate(arm64_exc_hwstate_t expected)
+{
+	arm64_exc_hwstate_t actual;
+
+	if (!IS_ENABLED(CONFIG_DEBUG_IRQFLAGS))
+		return;
+
+	actual.flags = arch_local_save_flags();
+
+	if (expected.flags == actual.flags)
+		return;
+
+	if (system_uses_irq_prio_masking()) {
+		WARN_ONCE(1, "Unexpected DAIF+PMR: 0x%x + 0x%x (expected 0x%x + 0x%x)\n",
+			  actual.daif, actual.pmr, expected.daif, expected.pmr);
+	} else {
+		WARN_ONCE(1, "Unexpected DAIF: 0x%x (expected 0x%x)\n",
+			  actual.daif, expected.daif);
+	}
+}
+
+static __always_inline
+void arm64_debug_exc_context(arm64_exc_context_t context)
+{
+	arm64_exc_hwstate_t expected = arm64_exc_hwstate_of_context(context);
+
+	arm64_debug_exc_hwstate(expected);
+}
+
+static __always_inline
+void arm64_update_exc_hwstate(arm64_exc_hwstate_t hwstate, bool update_pmr)
+{
+	if (system_uses_irq_prio_masking() &&
+	    update_pmr &&
+	    hwstate.pmr == GIC_PRIO_IRQOFF) {
+		/*
+		 * There has been concern that the write to daif
+		 * might be reordered before this write to PMR.
+		 * From the ARM ARM DDI 0487D.a, section D1.7.1
+		 * "Accessing PSTATE fields":
+		 *   Writes to the PSTATE fields have side-effects on
+		 *   various aspects of the PE operation. All of these
+		 *   side-effects are guaranteed:
+		 *     - Not to be visible to earlier instructions in
+		 *       the execution stream.
+		 *     - To be visible to later instructions in the
+		 *       execution stream
+		 *
+		 * Also, writes to PMR are self-synchronizing, so no
+		 * interrupts with a lower priority than PMR is signaled
+		 * to the PE after the write.
+		 *
+		 * So we don't need additional synchronization here.
+		 */
+		write_sysreg_s(hwstate.pmr, SYS_ICC_PMR_EL1);
+		pmr_sync();

According to the comment above, pmr_sync() seems unnecessary here, can
we safely remove it?

Yes.

quoted

+	}
+
+	write_sysreg(hwstate.daif, daif);
+
+	if (system_uses_irq_prio_masking() &&
+	    update_pmr &&
+	    hwstate.pmr == GIC_PRIO_IRQON) {
+		write_sysreg_s(hwstate.pmr, SYS_ICC_PMR_EL1);
+		pmr_sync();
+	}

I notice the sequence of manipulating DAIF and PMR for IRQ disabling
has been reordered from the old one, is there any subtle side effect
we should be aware of?

Here is my rationale for structuring it this way.

PMR only needs to be set to IRQOFF for NOIRQ context, so we can think
about the edge cases when entering and leaving NOIRQ context:

1. NONMI -> NOIRQ

   This means transitioning from IF+IRQON to 0+IRQOFF.

   The required order is:
     1. Set PMR to IRQOFF
     2. Clear IF

   Doing this in the reverse order would open a window for IRQs.

2. PROCESS -> NOIRQ

   This means transitioning from 0+IRQON to 0+IRQOFF.

   We do not care about DAIF here, so the only required step is:
     1. Set PMR to IRQOFF

3. NOIRQ -> NONMI

   This means transitioning from 0+IRQOFF to IF+IRQON.

   The required order is:
     1. Set DAIF to IF
     2. Set PMR to IRQON

   Doing this in the reverse order would open a window for IRQs.

4. NOIRQ -> PROCESS

   This means transitioning from 0+IRQOFF to 0+IRQON.

   We do not care about DAIF here, so the required order is:
     1. Set PMR to IRQON
     2. PMR sync

Notes:

  * Cases #1 and #3 also apply to ERROR and CRITICAL context; the only
    difference is the DAIF value.

  * Case #2 is equivalent to local_irq_disable().

  * Case #4 is equivalent to local_irq_enable().

So we can group the operations together as (pseudo-code):

  if IRQOFF
      Set PMR to IRQOFF

  Set DAIF to [value]

  if IRQON {
      Set PMR to IRQON;
      PMR sync;
  }

Additional notes:

* I agree that I was too defensive with pmr_sync() in the C code for the IRQOFF case.
* The C code also needs compiler barriers sprinkled in the right places.
* I will fix these in the next version.

Have I missed anything?

quoted

+}
+
+static __always_inline
+void arm64_update_exc_context(arm64_exc_context_t context, bool update_pmr)
+{
+	arm64_exc_hwstate_t hwstate = arm64_exc_hwstate_of_context(context);
+
+	arm64_update_exc_hwstate(hwstate, update_pmr);
+}
+#endif /* __ASM_INTERRUPTS_COMMON_FLAGS_H */

diff --git a/arch/arm64/include/asm/interrupts/entry.h b/arch/arm64/include/asm/interrupts/entry.h
new file mode 100644
index 000000000000..3034c490ed66
--- /dev/null
+++ b/arch/arm64/include/asm/interrupts/entry.h

@@ -0,0 +1,110 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (C) 2025 Arm Ltd.
+ */
+#ifndef __ASM_INTERRUPTS_ENTRY_H
+#define __ASM_INTERRUPTS_ENTRY_H
+
+#include <asm/arch_gicv3.h>
+#include <asm/bug.h>
+#include <asm/cpufeature.h>
+#include <asm/interrupts/common_flags.h>
+
+
+static __always_inline
+arm64_exc_hwstate_t __arm64_switch_exc_hwstate_to(arm64_exc_hwstate_t prev,
+						  arm64_exc_hwstate_t next)
+{
+	bool update_pmr = system_uses_irq_prio_masking() && prev.pmr != next.pmr;
+
+	arm64_debug_exc_hwstate(prev);
+
+	if (prev.flags == next.flags)
+		return next;
+
+	if (!arch_irqs_disabled_flags(next.flags))
+		trace_hardirqs_on();
+
+	arm64_update_exc_hwstate(next, update_pmr);
+
+	if (arch_irqs_disabled_flags(next.flags))

Could we save the disabled flags in a local variable to avoid the duplicate calls?

Yes we can, I will do it in the next version.

quoted

+		trace_hardirqs_off();
+
+	return next;
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_inherit_exc_context(struct pt_regs *regs)
+{
+	arm64_exc_hwstate_t prev = arm64_exc_hwstate_of_context(CRITICAL_CONTEXT);
+	arm64_exc_hwstate_t next = arm64_inherit_exc_hwstate(regs);
+
+	return __arm64_switch_exc_hwstate_to(prev, next);
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_drop_exc_context(arm64_exc_hwstate_t prev, arm64_exc_context_t context)
+{
+	arm64_exc_hwstate_t next = arm64_exc_hwstate_of_context(context);
+
+	if (IS_ENABLED(CONFIG_DEBUG_IRQFLAGS)) {
+		bool pnmi = system_uses_irq_prio_masking();
+
+		WARN_ON_ONCE(context > ERROR_CONTEXT &&
+			     prev.daif == DAIF_ERRCTX);
+
+		WARN_ON_ONCE(context > NONMI_CONTEXT &&
+			     prev.daif == DAIF_PROCCTX_NOIRQ);
+
+		WARN_ON_ONCE(context > NOIRQ_CONTEXT &&
+			     pnmi && prev.pmr == GIC_PRIO_IRQOFF);
+
+		WARN_ON_ONCE(context > PROCESS_CONTEXT &&
+			     ((pnmi && prev.daif == DAIF_PROCCTX && prev.pmr == GIC_PRIO_IRQON) ||
+			      (!pnmi && prev.daif == DAIF_PROCCTX)));
+	}
+
+	return __arm64_switch_exc_hwstate_to(prev, next);
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_lift_exc_context(arm64_exc_hwstate_t prev, arm64_exc_context_t context)
+{
+	arm64_exc_hwstate_t next = arm64_exc_hwstate_of_context(context);
+
+	if (IS_ENABLED(CONFIG_DEBUG_IRQFLAGS)) {
+		bool pnmi = system_uses_irq_prio_masking();
+
+		WARN_ON_ONCE(context < CRITICAL_CONTEXT &&
+			     prev.daif == DAIF_MASK);
+
+		WARN_ON_ONCE(context < ERROR_CONTEXT &&
+			     prev.daif == DAIF_ERRCTX);
+
+		WARN_ON_ONCE(context < NONMI_CONTEXT &&
+			     pnmi && prev.daif == DAIF_PROCCTX_NOIRQ);
+
+		WARN_ON_ONCE(context < NOIRQ_CONTEXT &&
+			     ((pnmi && prev.pmr == GIC_PRIO_IRQOFF) ||
+			      (!pnmi && prev.daif == DAIF_PROCCTX_NOIRQ)));
+	}
+
+	return __arm64_switch_exc_hwstate_to(prev, next);
+}
+
+
+static __always_inline
+arm64_exc_hwstate_t arm64_unmask_exc_context(arm64_exc_context_t context)
+{
+	arm64_exc_hwstate_t prev = arm64_exc_hwstate_of_context(CRITICAL_CONTEXT);
+
+	return arm64_drop_exc_context(prev, context);
+}
+
+static __always_inline
+arm64_exc_hwstate_t arm64_mask_exc_context(arm64_exc_hwstate_t prev)
+{
+	return arm64_lift_exc_context(prev, CRITICAL_CONTEXT);
+}
+
+#endif /* __ASM_INTERRUPTS_ENTRY_H */

-- BR Liao, Chang

Cheers
Vladimir