Do *not* delay while waiting for IOTLB invalidation to complete. A 1ms

delay is awful in a hot path, and the SMMU is actually quite quick on
invalidation, so simply removing the delay is worth a thousand roses.

Found with mental support from dlg@ (and btrace)

diff --git a/sys/arch/arm64/dev/smmu.c b/sys/arch/arm64/dev/smmu.c
index 5be6634..0b206a6 100644 (file)
--- a/sys/arch/arm64/dev/smmu.c
+++ b/sys/arch/arm64/dev/smmu.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: smmu.c,v 1.3 2021/03/01 21:38:20 patrick Exp $ */
+/* $OpenBSD: smmu.c,v 1.4 2021/03/02 01:34:43 patrick Exp $ */
 /*
  * Copyright (c) 2008-2009,2014-2016 Dale Rahn <drahn@dalerahn.com>
  * Copyright (c) 2021 Patrick Wildt <patrick@blueri.se>
@@ -389,7 +389,6 @@ smmu_tlb_sync_global(struct smmu_softc *sc)
                if ((smmu_gr0_read_4(sc, SMMU_STLBGSTATUS) &
                    SMMU_STLBGSTATUS_GSACTIVE) == 0)
                        return;
-               delay(1000);
        }
 
        printf("%s: global TLB sync timeout\n",
@@ -407,7 +406,6 @@ smmu_tlb_sync_context(struct smmu_domain *dom)
                if ((smmu_cb_read_4(sc, dom->sd_cb_idx, SMMU_CB_TLBSTATUS) &
                    SMMU_CB_TLBSTATUS_SACTIVE) == 0)
                        return;
-               delay(1000);
        }
 
        printf("%s: context TLB sync timeout\n",