Only do TLB shootdown on CPUs where a pmap is active. Only make SBI calls

that do a remote SFENCE.VMA or FENCE.I if the set of hart IDs isn't empty.
This significantly speeds things up.

ok jsg@

diff --git a/sys/arch/riscv64/riscv64/pmap.c b/sys/arch/riscv64/riscv64/pmap.c
index e8d263f..b2ab28c 100644 (file)
--- a/sys/arch/riscv64/riscv64/pmap.c
+++ b/sys/arch/riscv64/riscv64/pmap.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: pmap.c,v 1.15 2021/06/30 07:39:05 jsg Exp $   */
+/*     $OpenBSD: pmap.c,v 1.16 2021/07/02 08:53:28 kettenis Exp $      */
 
 /*
  * Copyright (c) 2019-2020 Brian Bamsch <bbamsch@google.com>
@@ -42,6 +42,16 @@ void pmap_free_asid(pmap_t);
 /* We run userland code with ASIDs that have the low bit set. */
 #define ASID_USER      1
 
+#ifdef MULTIPROCESSOR
+
+static inline int
+pmap_is_active(struct pmap *pm, struct cpu_info *ci)
+{
+       return pm == pmap_kernel() || pm == ci->ci_curpm;
+}
+
+#endif
+
 static inline void
 tlb_flush(pmap_t pm, vaddr_t va)
 {
@@ -53,10 +63,12 @@ tlb_flush(pmap_t pm, vaddr_t va)
        CPU_INFO_FOREACH(cii, ci) {
                if (ci == curcpu())
                        continue;
-               hart_mask |= (1UL << ci->ci_hartid);
+               if (pmap_is_active(pm, ci))
+                       hart_mask |= (1UL << ci->ci_hartid);
        }
 
-       sbi_remote_sfence_vma(&hart_mask, va, PAGE_SIZE);
+       if (hart_mask != 0)
+               sbi_remote_sfence_vma(&hart_mask, va, PAGE_SIZE);
 #endif
 
        if (pm == pmap_kernel()) {
@@ -83,7 +95,8 @@ icache_flush(void)
                hart_mask |= (1UL << ci->ci_hartid);
        }
 
-       sbi_remote_fence_i(&hart_mask);
+       if (hart_mask != 0)
+               sbi_remote_fence_i(&hart_mask);
 #endif
 
        fence_i();
@@ -2277,7 +2290,7 @@ pmap_set_satp(struct proc *p)
        struct cpu_info *ci = curcpu();
        pmap_t pm = p->p_vmspace->vm_map.pmap;
 
+       ci->ci_curpm = pm;
        load_satp(pm->pm_satp);
        __asm __volatile("sfence.vma");
-       ci->ci_curpm = pm;
 }