Linux-ARM-Kernel Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: "Kiryl Shutsemau (Meta)" <kas@kernel.org>
To: Vladimir Murzin <vladimir.murzin@arm.com>
Cc: mark.rutland@arm.com, catalin.marinas@arm.com,
	Jinjie Ruan <ruanjinjie@huawei.com>,
	liaochang1@huawei.com, ashok.raj@oss.qualcomm.com,
	maz@kernel.org, will@kernel.org, lpieralisi@kernel.org,
	linux-arm-kernel@lists.infradead.org
Subject: Re: [PATCH v2 00/39] arm64: Add support for FEAT_NMI
Date: Thu, 19 Nov 2026 16:38:51 +0000	[thread overview]
Message-ID: <av8lNnZPXOwlS7au@shutemov.name> (raw)
In-Reply-To: <av8bByU93_nCbPqE@shutemov.name>

On Thu, Nov 19, 2026 at 04:07:19PM +0000, Kiryl Shutsemau (Meta) wrote:
>   - Kernel: never signal a CPU that has SDEI masked. The signalling CPU
>     and the cpu_pm notifier can handshake: the sender marks the target
>     and IPIs it instead if it is already masked, and the notifier
>     refuses the power-down state while a signal is in flight. SDEI keeps
>     serving every running CPU, and the race goes away whatever the
>     firmware does. It needs a hook in arm_sdei.c.

Here is what I meant.

sdei_event_signal() refuses a masked target with -EAGAIN, and the
backtrace sends those CPUs the regular IPI. A PE that sees a signal on
its way stays out of the power-down state; the mask paths that cannot
fail wait for it, up to a millisecond.

On QEMU with TF-A minus the v2.13 check, masking a CPU through
cpu_pm_enter() and then signalling it gets -EAGAIN instead of the wedge,
and sysrq-l still goes over SDEI to the other CPUs.

I can send proper patches if it looks reasonable.

diff --git a/arch/arm64/include/asm/nmi.h b/arch/arm64/include/asm/nmi.h
index 2e8974ff8d63..561d062921e4 100644
--- a/arch/arm64/include/asm/nmi.h
+++ b/arch/arm64/include/asm/nmi.h
@@ -44,5 +44,6 @@ static inline void sdei_nmi_stop_cpus(const cpumask_t *mask) { }
  * Defined in arch/arm64/kernel/smp.c.
  */
 void __noreturn arm64_nmi_cpu_stop(struct pt_regs *regs, bool die_on_crash);
+void arm64_backtrace_ipi(cpumask_t *mask);
 
 #endif /* __ASM_NMI_H */
diff --git a/arch/arm64/kernel/smp.c b/arch/arm64/kernel/smp.c
index 014b163ce658..8369a4cd3e32 100644
--- a/arch/arm64/kernel/smp.c
+++ b/arch/arm64/kernel/smp.c
@@ -946,7 +946,7 @@ static void arm64_send_ipi(const cpumask_t *mask, unsigned int nr)
 			__ipi_send_single(get_ipi_desc(cpu, nr), cpu);
 }
 
-static void arm64_backtrace_ipi(cpumask_t *mask)
+void arm64_backtrace_ipi(cpumask_t *mask)
 {
 	arm64_send_ipi(mask, IPI_CPU_BACKTRACE);
 }
diff --git a/drivers/firmware/arm_sdei.c b/drivers/firmware/arm_sdei.c
index e8dd2f0f3919..86f4349fc364 100644
--- a/drivers/firmware/arm_sdei.c
+++ b/drivers/firmware/arm_sdei.c
@@ -12,6 +12,7 @@
 #include <linux/cpuhotplug.h>
 #include <linux/cpu.h>
 #include <linux/cpu_pm.h>
+#include <linux/delay.h>
 #include <linux/errno.h>
 #include <linux/hardirq.h>
 #include <linux/kernel.h>
@@ -32,6 +33,8 @@
 #include <linux/smp.h>
 #include <linux/spinlock.h>
 
+#include <asm/smp_plat.h>
+
 /*
  * The call to use to reach the firmware.
  */
@@ -299,10 +302,37 @@ static int sdei_api_get_version(u64 *version)
 	return invoke_sdei_fn(SDEI_1_0_FN_SDEI_VERSION, 0, 0, 0, 0, 0, version);
 }
 
+/*
+ * Event 0 must not land on a masked PE. The sender and the masking PE
+ * each set their own flag, then read the other's.
+ */
+static DEFINE_PER_CPU(bool, sdei_pe_unmasked);
+static DEFINE_PER_CPU(bool, sdei_signal_pending);
+
+static bool sdei_mark_masked(void)
+{
+	this_cpu_write(sdei_pe_unmasked, false);
+	smp_mb();
+	return this_cpu_read(sdei_signal_pending);
+}
+
+static void sdei_block_signal(void)
+{
+	int i;
+
+	if (!sdei_mark_masked())
+		return;
+
+	for (i = 0; i < USEC_PER_MSEC && this_cpu_read(sdei_signal_pending); i++)
+		udelay(1);
+}
+
 int sdei_mask_local_cpu(void)
 {
 	int err;
 
+	sdei_block_signal();
+
 	err = invoke_sdei_fn(SDEI_1_0_FN_SDEI_PE_MASK, 0, 0, 0, 0, 0, NULL);
 	if (err && err != -EIO) {
 		pr_warn_once("failed to mask CPU[%u]: %d\n",
@@ -330,6 +360,7 @@ int sdei_unmask_local_cpu(void)
 		return err;
 	}
 
+	this_cpu_write(sdei_pe_unmasked, true);
 	return 0;
 }
 
@@ -344,10 +375,23 @@ static void _ipi_unmask_cpu(void *ignored)
  * but the SMC -- no locks, no event lookup -- so it is safe from NMI /
  * crash context (e.g. the cross-CPU NMI service).
  */
-int sdei_event_signal(u32 event_num, u64 mpidr)
+int sdei_event_signal(u32 event_num, unsigned int cpu)
 {
-	return invoke_sdei_fn(SDEI_1_0_FN_SDEI_EVENT_SIGNAL, event_num,
-			      mpidr, 0, 0, 0, NULL);
+	int err;
+
+	WRITE_ONCE(per_cpu(sdei_signal_pending, cpu), true);
+	smp_mb();
+	if (!READ_ONCE(per_cpu(sdei_pe_unmasked, cpu))) {
+		WRITE_ONCE(per_cpu(sdei_signal_pending, cpu), false);
+		return -EAGAIN;
+	}
+
+	err = invoke_sdei_fn(SDEI_1_0_FN_SDEI_EVENT_SIGNAL, event_num,
+			     cpu_logical_map(cpu), 0, 0, 0, NULL);
+	if (err)
+		WRITE_ONCE(per_cpu(sdei_signal_pending, cpu), false);
+
+	return err;
 }
 NOKPROBE_SYMBOL(sdei_event_signal);
 
@@ -680,6 +724,8 @@ static int sdei_cpuhp_down(unsigned int cpu)
 	struct sdei_event *event;
 	int err;
 
+	sdei_block_signal();
+
 	/* un-register private events */
 	spin_lock(&sdei_list_lock);
 	list_for_each_entry(event, &sdei_list, list) {
@@ -702,6 +748,8 @@ static int sdei_cpuhp_up(unsigned int cpu)
 	struct sdei_event *event;
 	int err;
 
+	this_cpu_write(sdei_signal_pending, false);
+
 	/* re-register/enable private events */
 	spin_lock(&sdei_list_lock);
 	list_for_each_entry(event, &sdei_list, list) {
@@ -739,6 +787,10 @@ static int sdei_pm_notifier(struct notifier_block *nb, unsigned long action,
 
 	switch (action) {
 	case CPU_PM_ENTER:
+		if (sdei_mark_masked()) {
+			this_cpu_write(sdei_pe_unmasked, true);
+			return notifier_from_errno(-EBUSY);
+		}
 		rv = sdei_mask_local_cpu();
 		break;
 	case CPU_PM_EXIT:
@@ -1114,6 +1166,9 @@ int sdei_event_handler(struct pt_regs *regs,
 	int err;
 	u32 event_num = arg->event_num;
 
+	if (!event_num)
+		__this_cpu_write(sdei_signal_pending, false);
+
 	err = arg->callback(event_num, regs, arg->callback_arg);
 	if (err)
 		pr_err_ratelimited("event %u on CPU %u failed with error: %d\n",
diff --git a/drivers/firmware/arm_sdei_nmi.c b/drivers/firmware/arm_sdei_nmi.c
index 038eb79e28ac..9e2ea532888a 100644
--- a/drivers/firmware/arm_sdei_nmi.c
+++ b/drivers/firmware/arm_sdei_nmi.c
@@ -125,15 +125,19 @@ static int sdei_nmi_handler(u32 event, struct pt_regs *regs, void *arg)
 }
 NOKPROBE_SYMBOL(sdei_nmi_handler);
 
-static void sdei_nmi_fire(unsigned int target_cpu)
+static int sdei_nmi_fire(unsigned int target_cpu)
 {
-	int err = sdei_event_signal(SDEI_NMI_EVENT, cpu_logical_map(target_cpu));
+	int err = sdei_event_signal(SDEI_NMI_EVENT, target_cpu);
 
-	if (err)
+	if (err && err != -EAGAIN)
 		pr_warn("SDEI_EVENT_SIGNAL to CPU %u failed: %d\n",
 			target_cpu, err);
+
+	return err;
 }
 
+static cpumask_t sdei_nmi_ipi_mask;
+
 /*
  * Raise callback for nmi_trigger_cpumask_backtrace(): signal event 0
  * at every CPU still pending in @mask. The framework excludes the local
@@ -150,8 +154,14 @@ static void sdei_nmi_raise_backtrace(cpumask_t *mask)
 	 */
 	dsb(ishst);
 
-	for_each_cpu(cpu, mask)
-		sdei_nmi_fire(cpu);
+	cpumask_clear(&sdei_nmi_ipi_mask);
+	for_each_cpu(cpu, mask) {
+		if (sdei_nmi_fire(cpu) == -EAGAIN)
+			cpumask_set_cpu(cpu, &sdei_nmi_ipi_mask);
+	}
+
+	if (!cpumask_empty(&sdei_nmi_ipi_mask))
+		arm64_backtrace_ipi(&sdei_nmi_ipi_mask);
 }
 
 /*
diff --git a/include/linux/arm_sdei.h b/include/linux/arm_sdei.h
index b9dc21c241be..3f565a0f4a17 100644
--- a/include/linux/arm_sdei.h
+++ b/include/linux/arm_sdei.h
@@ -41,7 +41,7 @@ int sdei_event_disable(u32 event_num);
  * Signal the software-signalled event (event 0) to another PE, NMI-like.
  * @mpidr is the target's MPIDR affinity.
  */
-int sdei_event_signal(u32 event_num, u64 mpidr);
+int sdei_event_signal(u32 event_num, unsigned int cpu);
 
 /* Was SDEI firmware probed and usable? */
 bool sdei_is_present(void);
-- 
  Kiryl Shutsemau / Kirill A. Shutemov


  reply	other threads:[~2026-11-19 16:39 UTC|newest]

Thread overview: 57+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-12 13:20 [PATCH v2 00/39] arm64: Add support for FEAT_NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 01/39] arm64: irqflags: Introduce arm64-specific irqflags type Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 02/39] arm64: irqflags: save and use both DAIF and PMR Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 03/39] arm64: interrupts: Add common exception state helpers Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 04/39] arm64: process: Lift check for exception state to generic infrastructure Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 05/39] arm64: entry: Introduce entry specific exception masking helpers Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 06/39] arm64: entry: replace DAIF helpers with entry helpers Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 07/39] arm64: interrupts: Introduce exception masking save/restore helpers Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 08/39] arm64: interrupts: introduce a helper for GIC priority initialization Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 09/39] arm64: replace local_daif helpers Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 10/39] arm64: cpuidle: use new helpers to bypass interrupt priority masking Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 11/39] arm64: remove daifflags.h Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 12/39] arm64: gicv3: remove GIC_PRIO_PSR_I_SET Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 13/39] arm64: cpufeature: Remove system_has_prio_mask_debugging() Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 14/39] arm64: irqflags: Switch to CONFIG_DEBUG_IRQFLAGS Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 15/39] arm64: Kconfig: Remove CONFIG_ARM64_DEBUG_PRIORITY_MASKING Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 16/39] efi/runtime-wrappers: Permit architectures to override IRQ flags checks Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 17/39] arm64/efi: Implement override for " Vladimir Murzin
2026-11-13 13:22   ` Mark Brown
2026-11-12 13:20 ` [PATCH v2 18/39] arm64: booting: Document boot requirements for FEAT_NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 19/39] arm64: sysreg: Add definitions for immediate versions of MSR ALLINT Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 20/39] arm64: probes: Reject stepping over ALLINT Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 21/39] arm64: ptrace: Add PSR_ALLINT_BIT Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 22/39] arm64: idreg: Add an override for FEAT_NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 23/39] arm64: cpufeature: Detect PE support " Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 24/39] arm64: nmi: Manage masking for superpriority interrupts Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 25/39] arm64: irq: Report FEAT_NMI masking local IRQs Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 26/39] arm64: nmi: Add handling of superpriority interrupts as NMIs Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 27/39] arm64: suspend: Always initialise PSTATE.ALLINT Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 28/39] arm64/efi: Add ALLINT to IRQ flags checks Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 29/39] arm64: kprobes: Disable NMIs Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 30/39] arm64: smp: Abstract SGI and LPI operations Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 31/39] arm64: smp: Fall back to IRQ when IPI NMI request fails Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 32/39] arm64: nmi: Add Kconfig for NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 33/39] arm64: nmi: Select HAVE_PERF_EVENTS_NMI for FEAT_NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 34/39] irqchip/gic-v3: Rename nmi to pseudo_nmi Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 35/39] irqchip/gic-v3: Implement FEAT_GICv3_NMI support Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 36/39] arm64: smp: Add NMI support for LPI-backed IPIs Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 37/39] genirq: Trust local irqchip when considering IRQ for NMI Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 38/39] genirq: Do not call add_interrupt_randomness() in NMI context Vladimir Murzin
2026-11-12 13:20 ` [PATCH v2 39/39] irqchip/gic-v5: Add NMI support Vladimir Murzin
2026-11-14  5:03 ` [PATCH v2 00/39] arm64: Add support for FEAT_NMI Ashok Raj
2026-11-19 11:18   ` Vladimir Murzin
2026-11-16  6:58 ` Jinjie Ruan
2026-11-16 10:02   ` Vladimir Murzin
2026-11-16 11:01     ` Jinjie Ruan
2026-11-18 13:22       ` Vladimir Murzin
2026-11-19  2:33         ` Jinjie Ruan
2026-11-18 12:17 ` Jinjie Ruan
2026-11-19  1:54   ` Jinjie Ruan
2026-11-19 10:14     ` Vladimir Murzin
2026-11-19 12:45       ` Jinjie Ruan
2026-11-19 14:50         ` Vladimir Murzin
2026-11-19 16:07           ` Kiryl Shutsemau (Meta)
2026-11-19 16:38             ` Kiryl Shutsemau (Meta) [this message]
2026-11-19 11:41   ` Jinjie Ruan
2026-11-19 14:46     ` Vladimir Murzin

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=av8lNnZPXOwlS7au@shutemov.name \
    --to=kas@kernel.org \
    --cc=ashok.raj@oss.qualcomm.com \
    --cc=catalin.marinas@arm.com \
    --cc=liaochang1@huawei.com \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=lpieralisi@kernel.org \
    --cc=mark.rutland@arm.com \
    --cc=maz@kernel.org \
    --cc=ruanjinjie@huawei.com \
    --cc=vladimir.murzin@arm.com \
    --cc=will@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox