diff --git a/sys/kern/pic_if.m b/sys/kern/pic_if.m
--- a/sys/kern/pic_if.m
+++ b/sys/kern/pic_if.m
@@ -89,6 +89,14 @@
 
 		return (EOPNOTSUPP);
 	}
+
+	static int
+	dflt_pic_ipi_setup_per_cpu(device_t dev, u_int ipi,
+	    struct intr_irqsrc *isrc, u_int cpu)
+	{
+
+		return (EOPNOTSUPP);
+	}
 };
 
 METHOD int activate_intr {
@@ -172,3 +180,10 @@
 	u_int			ipi;
 	struct intr_irqsrc	**isrcp;
 } DEFAULT dflt_pic_ipi_setup;
+
+METHOD int ipi_setup_per_cpu {
+	device_t		dev;
+	u_int			ipi;
+	struct intr_irqsrc	**isrcp;
+	u_int			cpu;
+} DEFAULT dflt_pic_ipi_setup_per_cpu;
diff --git a/sys/kern/subr_intr.c b/sys/kern/subr_intr.c
--- a/sys/kern/subr_intr.c
+++ b/sys/kern/subr_intr.c
@@ -137,13 +137,18 @@
 struct intr_ipi {
 	intr_ipi_handler_t	*ii_handler;
 	void			*ii_handler_arg;
-	struct intr_irqsrc	*ii_isrc;
+	union {
+		struct intr_irqsrc	**ii_isrc_per_cpu;
+		struct intr_irqsrc	*ii_isrc;
+	};
 	char			ii_name[INTR_IPI_NAMELEN];
 	u_long			*ii_count;
+	bool			ii_isrc_is_per_cpu;
 };
 
 static device_t intr_ipi_dev;
 static u_int intr_ipi_dev_priority;
+static uint32_t intr_ipi_flags;
 static bool intr_ipi_dev_frozen;
 #endif
 
@@ -1840,7 +1845,7 @@
 }
 
 int
-intr_ipi_pic_register(device_t dev, u_int priority)
+intr_ipi_pic_register_flags(device_t dev, u_int priority, uint32_t flags)
 {
 	if (intr_ipi_dev_frozen) {
 		device_printf(dev, "IPI device already frozen");
@@ -1850,11 +1855,18 @@
 	if (intr_ipi_dev == NULL || priority > intr_ipi_dev_priority) {
 		intr_ipi_dev_priority = priority;
 		intr_ipi_dev = dev;
+		intr_ipi_flags = flags;
 	}
 
 	return (0);
 }
 
+int
+intr_ipi_pic_register(device_t dev, u_int priority)
+{
+	return (intr_ipi_pic_register_flags(dev, priority, 0));
+}
+
 /*
  *  Setup IPI handler on interrupt controller.
  *
@@ -1866,7 +1878,7 @@
 {
 	struct intr_irqsrc *isrc;
 	struct intr_ipi *ii;
-	int error;
+	int cpu, error;
 
 	if (!intr_ipi_dev_frozen) {
 		if (intr_ipi_dev == NULL)
@@ -1878,22 +1890,52 @@
 
 	KASSERT(hand != NULL, ("%s: ipi %u no handler", __func__, ipi));
 
-	error = PIC_IPI_SETUP(intr_ipi_dev, ipi, &isrc);
-	if (error != 0)
-		return;
-
-	isrc->isrc_handlers++;
-
 	ii = intr_ipi_lookup(ipi);
 	KASSERT(ii->ii_count == NULL, ("%s: ipi %u reused", __func__, ipi));
 
+	ii->ii_isrc_is_per_cpu = (intr_ipi_flags & INTR_IPI_PER_CPU) != 0;
+
+	if (!ii->ii_isrc_is_per_cpu) {
+		error = PIC_IPI_SETUP(intr_ipi_dev, ipi, &isrc);
+		if (error != 0)
+			return;
+
+		ii->ii_isrc = isrc;
+		isrc->isrc_handlers++;
+	} else {
+		ii->ii_isrc_per_cpu = mallocarray(mp_maxid + 1,
+		    sizeof(*ii->ii_isrc_per_cpu), M_INTRNG, M_WAITOK | M_ZERO);
+
+		/*
+		 * all_cpus may not have been initialized when this is called
+		 * so CPU_FOREACH won't work.
+		 */
+		for (cpu = 0; cpu <= mp_maxid; cpu++) {
+			error = PIC_IPI_SETUP_PER_CPU(intr_ipi_dev, ipi, &isrc,
+			    cpu);
+			if (error != 0) {
+				free(ii->ii_isrc_per_cpu, M_INTRNG);
+				ii->ii_isrc_per_cpu = NULL;
+				return;
+			}
+
+			ii->ii_isrc_per_cpu[cpu] = isrc;
+			isrc->isrc_handlers++;
+		}
+	}
+
 	ii->ii_handler = hand;
 	ii->ii_handler_arg = arg;
-	ii->ii_isrc = isrc;
 	strlcpy(ii->ii_name, name, INTR_IPI_NAMELEN);
 	ii->ii_count = intr_ipi_setup_counters(name);
 
-	PIC_ENABLE_INTR(intr_ipi_dev, isrc);
+	if (!ii->ii_isrc_is_per_cpu) {
+		PIC_ENABLE_INTR(intr_ipi_dev, ii->ii_isrc);
+	} else {
+		for (cpu = 0; cpu <= mp_maxid; cpu++) {
+			PIC_ENABLE_INTR(intr_ipi_dev, ii->ii_isrc_per_cpu[cpu]);
+		}
+	}
 }
 
 void
@@ -1921,7 +1963,16 @@
 	dsb(ishst);
 #endif
 
-	PIC_IPI_SEND(intr_ipi_dev, ii->ii_isrc, cpus, ipi);
+	if (ii->ii_isrc_is_per_cpu) {
+		u_int cpu;
+
+		CPU_FOREACH_ISSET(cpu, &cpus) {
+			PIC_IPI_SEND(intr_ipi_dev, ii->ii_isrc_per_cpu[cpu],
+			    cpus, ipi);
+		}
+	} else {
+		PIC_IPI_SEND(intr_ipi_dev, ii->ii_isrc, cpus, ipi);
+	}
 }
 
 /*
diff --git a/sys/sys/intr.h b/sys/sys/intr.h
--- a/sys/sys/intr.h
+++ b/sys/sys/intr.h
@@ -159,6 +159,10 @@
 #ifdef SMP
 typedef void intr_ipi_handler_t(void *);
 
+/* Flags for intr_ipi_pic_register_flags */
+#define	INTR_IPI_PER_CPU		(0x1u << 0)
+
+int intr_ipi_pic_register_flags(device_t dev, u_int priority, uint32_t flags);
 int intr_ipi_pic_register(device_t dev, u_int priority);
 void intr_ipi_setup(u_int ipi, const char *name, intr_ipi_handler_t *hand,
     void *arg);