Index: sys/amd64/amd64/mp_machdep.c
===================================================================
--- sys/amd64/amd64/mp_machdep.c
+++ sys/amd64/amd64/mp_machdep.c
@@ -638,8 +638,8 @@
  * Used by pmap to request cache or TLB invalidation on local and
  * remote processors.  Mask provides the set of remote CPUs which are
  * to be signalled with the invalidation IPI.  As an optimization, the
- * curcpu_cb callback is invoked on the calling CPU while waiting for
- * remote CPUs to complete the operation.
+ * curcpu_cb callback is invoked on the calling CPU in critical
+ * section while waiting for remote CPUs to complete the operation.
  *
  * The callback function is called unconditionally on the caller's
  * underlying processor, even when this processor is not set in the
@@ -662,23 +662,19 @@
 	 * It is not necessary to signal other CPUs while booting or
 	 * when in the debugger.
 	 */
-	if (kdb_active || KERNEL_PANICKED() || !smp_started) {
-		curcpu_cb(pmap, addr1, addr2);
-		return;
-	}
-
-	sched_pin();
+	if (kdb_active || KERNEL_PANICKED() || !smp_started)
+		goto local_cb;
 
 	/*
 	 * Check for other cpus.  Return if none.
 	 */
 	if (CPU_ISFULLSET(&mask)) {
 		if (mp_ncpus <= 1)
-			goto nospinexit;
+			goto local_cb;
 	} else {
 		CPU_CLR(PCPU_GET(cpuid), &mask);
 		if (CPU_EMPTY(&mask))
-			goto nospinexit;
+			goto local_cb;
 	}
 
 	/*
@@ -735,12 +731,12 @@
 			ia32_pause();
 	}
 	critical_exit();
-	sched_unpin();
 	return;
 
-nospinexit:
+local_cb:
+	critical_enter();
 	curcpu_cb(pmap, addr1, addr2);
-	sched_unpin();
+	critical_exit();
 }
 
 void
Index: sys/amd64/amd64/pmap.c
===================================================================
--- sys/amd64/amd64/pmap.c
+++ sys/amd64/amd64/pmap.c
@@ -2773,53 +2773,16 @@
 static cpuset_t
 pmap_invalidate_cpu_mask(pmap_t pmap)
 {
-
 	return (pmap == kernel_pmap ? all_cpus : pmap->pm_active);
 }
 
 static inline void
-pmap_invalidate_page_pcid(pmap_t pmap, vm_offset_t va,
-    const bool invpcid_works1)
+pmap_invalidate_preipi_pcid(pmap_t pmap)
 {
-	struct invpcid_descr d;
-	uint64_t kcr3, ucr3;
-	uint32_t pcid;
 	u_int cpuid, i;
 
 	cpuid = PCPU_GET(cpuid);
-	if (pmap == PCPU_GET(curpmap)) {
-		if (pmap->pm_ucr3 != PMAP_NO_CR3 &&
-		    /*
-		     * If we context-switched right after
-		     * PCPU_GET(ucr3_load_mask), we could read the
-		     * ~CR3_PCID_SAVE mask, which causes us to skip
-		     * the code below to invalidate user pages.  This
-		     * is handled in pmap_activate_sw_pcid_pti() by
-		     * clearing pm_gen if ucr3_load_mask is ~CR3_PCID_SAVE.
-		     */
-		    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
-			/*
-			 * Because pm_pcid is recalculated on a
-			 * context switch, we must disable switching.
-			 * Otherwise, we might use a stale value
-			 * below.
-			 */
-			critical_enter();
-			pcid = pmap->pm_pcids[cpuid].pm_pcid;
-			if (invpcid_works1) {
-				d.pcid = pcid | PMAP_PCID_USER_PT;
-				d.pad = 0;
-				d.addr = va;
-				invpcid(&d, INVPCID_ADDR);
-			} else {
-				kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
-				ucr3 = pmap->pm_ucr3 | pcid |
-				    PMAP_PCID_USER_PT | CR3_PCID_SAVE;
-				pmap_pti_pcid_invlpg(ucr3, kcr3, va);
-			}
-			critical_exit();
-		}
-	} else
+	if (pmap != PCPU_GET(curpmap))
 		pmap->pm_pcids[cpuid].pm_gen = 0;
 
 	CPU_FOREACH(i) {
@@ -2841,51 +2804,97 @@
 }
 
 static void
-pmap_invalidate_page_pcid_invpcid(pmap_t pmap, vm_offset_t va)
+pmap_invalidate_preipi_nopcid(pmap_t pmap __unused)
 {
+}
 
-	pmap_invalidate_page_pcid(pmap, va, true);
+DEFINE_IFUNC(static, void, pmap_invalidate_preipi, (pmap_t))
+{
+	return (pmap_pcid_enabled ? pmap_invalidate_preipi_pcid :
+	    pmap_invalidate_preipi_nopcid);
 }
 
-static void
-pmap_invalidate_page_pcid_noinvpcid(pmap_t pmap, vm_offset_t va)
+static inline void
+pmap_invalidate_page_pcid_cb(pmap_t pmap, vm_offset_t va,
+    const bool invpcid_works1)
 {
+	struct invpcid_descr d;
+	uint64_t kcr3, ucr3;
+	uint32_t pcid;
+	u_int cpuid;
+
+	/*
+	 * If we context-switched right after
+	 * PCPU_GET(ucr3_load_mask), we could read the ~CR3_PCID_SAVE
+	 * mask, which causes us to skip the code below to invalidate
+	 * user pages.  This is handled in pmap_activate_sw_pcid_pti()
+	 * by clearing pm_gen if ucr3_load_mask is ~CR3_PCID_SAVE.
+	 */
+	if (pmap->pm_ucr3 == PMAP_NO_CR3 ||
+	    PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK)
+		return;
+
+	/*
+	 * Because pm_pcid is recalculated on a context switch, we
+	 * must ensure there is no switching, not just pinning.
+	 * Otherwise, we might use a stale value below.
+	 */
+	CRITICAL_ASSERT(curthread);
+	cpuid = PCPU_GET(cpuid);
 
-	pmap_invalidate_page_pcid(pmap, va, false);
+	pcid = pmap->pm_pcids[cpuid].pm_pcid;
+	if (invpcid_works1) {
+		d.pcid = pcid | PMAP_PCID_USER_PT;
+		d.pad = 0;
+		d.addr = va;
+		invpcid(&d, INVPCID_ADDR);
+	} else {
+		kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
+		ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
+		pmap_pti_pcid_invlpg(ucr3, kcr3, va);
+	}
+}
+
+static void
+pmap_invalidate_page_pcid_invpcid_cb(pmap_t pmap, vm_offset_t va)
+{
+	pmap_invalidate_page_pcid_cb(pmap, va, true);
 }
 
 static void
-pmap_invalidate_page_nopcid(pmap_t pmap, vm_offset_t va)
+pmap_invalidate_page_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t va)
 {
+	pmap_invalidate_page_pcid_cb(pmap, va, false);
 }
 
-DEFINE_IFUNC(static, void, pmap_invalidate_page_mode, (pmap_t, vm_offset_t))
+static void
+pmap_invalidate_page_nopcid_cb(pmap_t pmap __unused, vm_offset_t va __unused)
 {
+}
 
+DEFINE_IFUNC(static, void, pmap_invalidate_page_cb, (pmap_t, vm_offset_t))
+{
 	if (pmap_pcid_enabled)
-		return (invpcid_works ? pmap_invalidate_page_pcid_invpcid :
-		    pmap_invalidate_page_pcid_noinvpcid);
-	return (pmap_invalidate_page_nopcid);
+		return (invpcid_works ? pmap_invalidate_page_pcid_invpcid_cb :
+		    pmap_invalidate_page_pcid_noinvpcid_cb);
+	return (pmap_invalidate_page_nopcid_cb);
 }
 
 static void
 pmap_invalidate_page_curcpu_cb(pmap_t pmap, vm_offset_t va,
     vm_offset_t addr2 __unused)
 {
-
 	if (pmap == kernel_pmap) {
 		invlpg(va);
-	} else {
-		if (pmap == PCPU_GET(curpmap))
-			invlpg(va);
-		pmap_invalidate_page_mode(pmap, va);
+	} else if (pmap == PCPU_GET(curpmap)) {
+		invlpg(va);
+		pmap_invalidate_page_cb(pmap, va);
 	}
 }
 
 void
 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
 {
-
 	if (pmap_type_guest(pmap)) {
 		pmap_invalidate_ept(pmap);
 		return;
@@ -2894,82 +2903,73 @@
 	KASSERT(pmap->pm_type == PT_X86,
 	    ("pmap_invalidate_page: invalid type %d", pmap->pm_type));
 
+	sched_pin();
+	pmap_invalidate_preipi(pmap);
 	smp_masked_invlpg(pmap_invalidate_cpu_mask(pmap), va, pmap,
 	    pmap_invalidate_page_curcpu_cb);
+	sched_unpin();
 }
 
 /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */
 #define	PMAP_INVLPG_THRESHOLD	(4 * 1024 * PAGE_SIZE)
 
 static void
-pmap_invalidate_range_pcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
+pmap_invalidate_range_pcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva,
     const bool invpcid_works1)
 {
 	struct invpcid_descr d;
 	uint64_t kcr3, ucr3;
 	uint32_t pcid;
-	u_int cpuid, i;
+	u_int cpuid;
+
+	if (pmap != PCPU_GET(curpmap) ||
+	    pmap->pm_ucr3 == PMAP_NO_CR3 ||
+	    PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK)
+		return;
 
+	CRITICAL_ASSERT(curthread);
 	cpuid = PCPU_GET(cpuid);
-	if (pmap == PCPU_GET(curpmap)) {
-		if (pmap->pm_ucr3 != PMAP_NO_CR3 &&
-		    PCPU_GET(ucr3_load_mask) == PMAP_UCR3_NOMASK) {
-			critical_enter();
-			pcid = pmap->pm_pcids[cpuid].pm_pcid;
-			if (invpcid_works1) {
-				d.pcid = pcid | PMAP_PCID_USER_PT;
-				d.pad = 0;
-				d.addr = sva;
-				for (; d.addr < eva; d.addr += PAGE_SIZE)
-					invpcid(&d, INVPCID_ADDR);
-			} else {
-				kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
-				ucr3 = pmap->pm_ucr3 | pcid |
-				    PMAP_PCID_USER_PT | CR3_PCID_SAVE;
-				pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
-			}
-			critical_exit();
-		}
-	} else
-		pmap->pm_pcids[cpuid].pm_gen = 0;
 
-	CPU_FOREACH(i) {
-		if (cpuid != i)
-			pmap->pm_pcids[i].pm_gen = 0;
+	pcid = pmap->pm_pcids[cpuid].pm_pcid;
+	if (invpcid_works1) {
+		d.pcid = pcid | PMAP_PCID_USER_PT;
+		d.pad = 0;
+		for (d.addr = sva; d.addr < eva; d.addr += PAGE_SIZE)
+			invpcid(&d, INVPCID_ADDR);
+	} else {
+		kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE;
+		ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE;
+		pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva);
 	}
-	/* See the comment in pmap_invalidate_page_pcid(). */
-	atomic_thread_fence_seq_cst();
 }
 
 static void
-pmap_invalidate_range_pcid_invpcid(pmap_t pmap, vm_offset_t sva,
+pmap_invalidate_range_pcid_invpcid_cb(pmap_t pmap, vm_offset_t sva,
     vm_offset_t eva)
 {
-
-	pmap_invalidate_range_pcid(pmap, sva, eva, true);
+	pmap_invalidate_range_pcid_cb(pmap, sva, eva, true);
 }
 
 static void
-pmap_invalidate_range_pcid_noinvpcid(pmap_t pmap, vm_offset_t sva,
+pmap_invalidate_range_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t sva,
     vm_offset_t eva)
 {
-
-	pmap_invalidate_range_pcid(pmap, sva, eva, false);
+	pmap_invalidate_range_pcid_cb(pmap, sva, eva, false);
 }
 
 static void
-pmap_invalidate_range_nopcid(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
+pmap_invalidate_range_nopcid_cb(pmap_t pmap __unused, vm_offset_t sva __unused,
+    vm_offset_t eva __unused)
 {
 }
 
-DEFINE_IFUNC(static, void, pmap_invalidate_range_mode, (pmap_t, vm_offset_t,
+DEFINE_IFUNC(static, void, pmap_invalidate_range_cb, (pmap_t, vm_offset_t,
     vm_offset_t))
 {
-
 	if (pmap_pcid_enabled)
-		return (invpcid_works ? pmap_invalidate_range_pcid_invpcid :
-		    pmap_invalidate_range_pcid_noinvpcid);
-	return (pmap_invalidate_range_nopcid);
+		return (invpcid_works ? pmap_invalidate_range_pcid_invpcid_cb :
+		    pmap_invalidate_range_pcid_noinvpcid_cb);
+	return (pmap_invalidate_range_nopcid_cb);
 }
 
 static void
@@ -2980,19 +2980,16 @@
 	if (pmap == kernel_pmap) {
 		for (addr = sva; addr < eva; addr += PAGE_SIZE)
 			invlpg(addr);
-	} else {
-		if (pmap == PCPU_GET(curpmap)) {
-			for (addr = sva; addr < eva; addr += PAGE_SIZE)
-				invlpg(addr);
-		}
-		pmap_invalidate_range_mode(pmap, sva, eva);
+	} else if (pmap == PCPU_GET(curpmap)) {
+		for (addr = sva; addr < eva; addr += PAGE_SIZE)
+			invlpg(addr);
+		pmap_invalidate_range_cb(pmap, sva, eva);
 	}
 }
 
 void
 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
 {
-
 	if (eva - sva >= PMAP_INVLPG_THRESHOLD) {
 		pmap_invalidate_all(pmap);
 		return;
@@ -3006,17 +3003,20 @@
 	KASSERT(pmap->pm_type == PT_X86,
 	    ("pmap_invalidate_range: invalid type %d", pmap->pm_type));
 
+	sched_pin();
+	pmap_invalidate_preipi(pmap);
 	smp_masked_invlpg_range(pmap_invalidate_cpu_mask(pmap), sva, eva, pmap,
 	    pmap_invalidate_range_curcpu_cb);
+	sched_unpin();
 }
 
 static inline void
-pmap_invalidate_all_pcid(pmap_t pmap, bool invpcid_works1)
+pmap_invalidate_all_pcid_cb(pmap_t pmap, bool invpcid_works1)
 {
 	struct invpcid_descr d;
 	uint64_t kcr3;
 	uint32_t pcid;
-	u_int cpuid, i;
+	u_int cpuid;
 
 	if (pmap == kernel_pmap) {
 		if (invpcid_works1) {
@@ -3025,79 +3025,64 @@
 		} else {
 			invltlb_glob();
 		}
-	} else {
+	} else if (pmap == PCPU_GET(curpmap)) {
+		CRITICAL_ASSERT(curthread);
 		cpuid = PCPU_GET(cpuid);
-		if (pmap == PCPU_GET(curpmap)) {
-			critical_enter();
-			pcid = pmap->pm_pcids[cpuid].pm_pcid;
-			if (invpcid_works1) {
-				d.pcid = pcid;
-				d.pad = 0;
-				d.addr = 0;
-				invpcid(&d, INVPCID_CTX);
-			} else {
-				kcr3 = pmap->pm_cr3 | pcid;
-				load_cr3(kcr3);
-			}
-			if (pmap->pm_ucr3 != PMAP_NO_CR3)
-				PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
-			critical_exit();
-		} else
-			pmap->pm_pcids[cpuid].pm_gen = 0;
-		CPU_FOREACH(i) {
-			if (cpuid != i)
-				pmap->pm_pcids[i].pm_gen = 0;
+
+		pcid = pmap->pm_pcids[cpuid].pm_pcid;
+		if (invpcid_works1) {
+			d.pcid = pcid;
+			d.pad = 0;
+			d.addr = 0;
+			invpcid(&d, INVPCID_CTX);
+		} else {
+			kcr3 = pmap->pm_cr3 | pcid;
+			load_cr3(kcr3);
 		}
+		if (pmap->pm_ucr3 != PMAP_NO_CR3)
+			PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE);
 	}
-	/* See the comment in pmap_invalidate_page_pcid(). */
-	atomic_thread_fence_seq_cst();
 }
 
 static void
-pmap_invalidate_all_pcid_invpcid(pmap_t pmap)
+pmap_invalidate_all_pcid_invpcid_cb(pmap_t pmap)
 {
-
-	pmap_invalidate_all_pcid(pmap, true);
+	pmap_invalidate_all_pcid_cb(pmap, true);
 }
 
 static void
-pmap_invalidate_all_pcid_noinvpcid(pmap_t pmap)
+pmap_invalidate_all_pcid_noinvpcid_cb(pmap_t pmap)
 {
-
-	pmap_invalidate_all_pcid(pmap, false);
+	pmap_invalidate_all_pcid_cb(pmap, false);
 }
 
 static void
-pmap_invalidate_all_nopcid(pmap_t pmap)
+pmap_invalidate_all_nopcid_cb(pmap_t pmap)
 {
-
 	if (pmap == kernel_pmap)
 		invltlb_glob();
 	else if (pmap == PCPU_GET(curpmap))
 		invltlb();
 }
 
-DEFINE_IFUNC(static, void, pmap_invalidate_all_mode, (pmap_t))
+DEFINE_IFUNC(static, void, pmap_invalidate_all_cb, (pmap_t))
 {
-
 	if (pmap_pcid_enabled)
-		return (invpcid_works ? pmap_invalidate_all_pcid_invpcid :
-		    pmap_invalidate_all_pcid_noinvpcid);
-	return (pmap_invalidate_all_nopcid);
+		return (invpcid_works ? pmap_invalidate_all_pcid_invpcid_cb :
+		    pmap_invalidate_all_pcid_noinvpcid_cb);
+	return (pmap_invalidate_all_nopcid_cb);
 }
 
 static void
 pmap_invalidate_all_curcpu_cb(pmap_t pmap, vm_offset_t addr1 __unused,
     vm_offset_t addr2 __unused)
 {
-
-	pmap_invalidate_all_mode(pmap);
+	pmap_invalidate_all_cb(pmap);
 }
 
 void
 pmap_invalidate_all(pmap_t pmap)
 {
-
 	if (pmap_type_guest(pmap)) {
 		pmap_invalidate_ept(pmap);
 		return;
@@ -3106,23 +3091,26 @@
 	KASSERT(pmap->pm_type == PT_X86,
 	    ("pmap_invalidate_all: invalid type %d", pmap->pm_type));
 
+	sched_pin();
+	pmap_invalidate_preipi(pmap);
 	smp_masked_invltlb(pmap_invalidate_cpu_mask(pmap), pmap,
 	    pmap_invalidate_all_curcpu_cb);
+	sched_unpin();
 }
 
 static void
 pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused, vm_offset_t va __unused,
     vm_offset_t addr2 __unused)
 {
-
 	wbinvd();
 }
 
 void
 pmap_invalidate_cache(void)
 {
-
+	sched_pin();
 	smp_cache_flush(pmap_invalidate_cache_curcpu_cb);
+	sched_unpin();
 }
 
 struct pde_action {