Blob Blame Raw
From eb121ffa97c1c25d7853d51b4c8209c0bb521deb Mon Sep 17 00:00:00 2001
From: David Gibson <dgibson@redhat.com>
Date: Fri, 7 Feb 2020 00:57:04 +0000
Subject: [PATCH 1/7] spapr: Enable DD2.3 accelerated count cache flush in
 pseries-5.0 machine

RH-Author: David Gibson <dgibson@redhat.com>
Message-id: <20200207005704.194428-1-dgibson@redhat.com>
Patchwork-id: 93737
O-Subject: [RHEL-AV-8.2 qemu-kvm PATCHv2] spapr: Enable DD2.3 accelerated count cache flush in pseries-5.0 machine
Bugzilla: 1796240
RH-Acked-by: Danilo de Paula <ddepaula@redhat.com>
RH-Acked-by: Laurent Vivier <lvivier@redhat.com>
RH-Acked-by: Thomas Huth <thuth@redhat.com>

From: David Gibson <david@gibson.dropbear.id.au>

For POWER9 DD2.2 cpus, the best current Spectre v2 indirect branch
mitigation is "count cache disabled", which is configured with:
    -machine cap-ibs=fixed-ccd
However, this option isn't available on DD2.3 CPUs with KVM, because they
don't have the count cache disabled.

For POWER9 DD2.3 cpus, it is "count cache flush with assist", configured
with:
    -machine cap-ibs=workaround,cap-ccf-assist=on
However this option isn't available on DD2.2 CPUs with KVM, because they
don't have the special CCF assist instruction this relies on.

On current machine types, we default to "count cache flush w/o assist",
that is:
    -machine cap-ibs=workaround,cap-ccf-assist=off
This runs, with mitigation on both DD2.2 and DD2.3 host cpus, but has a
fairly significant performance impact.

It turns out we can do better.  The special instruction that CCF assist
uses to trigger a count cache flush is a no-op on earlier CPUs, rather than
trapping or causing other badness.  It doesn't, of itself, implement the
mitigation, but *if* we have count-cache-disabled, then the count cache
flush is unnecessary, and so using the count cache flush mitigation is
harmless.

Therefore for the new pseries-5.0 machine type, enable cap-ccf-assist by
default.  Along with that, suppress throwing an error if cap-ccf-assist
is selected but KVM doesn't support it, as long as KVM *is* giving us
count-cache-disabled.  To allow TCG to work out of the box, even though it
doesn't implement the ccf flush assist, downgrade the error in that case to
a warning.  This matches several Spectre mitigations where we allow TCG
to operate for debugging, since we don't really make guarantees about TCG
security properties anyway.

While we're there, make the TCG warning for this case match that for other
mitigations.

Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Tested-by: Michael Ellerman <mpe@ellerman.id.au>
(cherry picked from commit 37965dfe4dffa3ac49438337417608e7f346b58a)
Signed-off-by: Danilo C. L. de Paula <ddepaula@redhat.com>

Conflicts:
	hw/ppc/spapr.c

Adjusted machine version compatibility code to the RHEL machine types
rather than the upstream machine types.

Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1796240
Brew: https://brewweb.engineering.redhat.com/brew/taskinfo?taskID=26285002
Branch: rhel-av-8.2.0
Upstream: Merged for qemu-5.0

Signed-off-by: David Gibson <dgibson@redhat.com>
Signed-off-by: Danilo C. L. de Paula <ddepaula@redhat.com>
---
 hw/ppc/spapr.c      |  4 +++-
 hw/ppc/spapr_caps.c | 21 +++++++++++++++++----
 2 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/hw/ppc/spapr.c b/hw/ppc/spapr.c
index c12862d..a330f03 100644
--- a/hw/ppc/spapr.c
+++ b/hw/ppc/spapr.c
@@ -4440,7 +4440,7 @@ static void spapr_machine_class_init(ObjectClass *oc, void *data)
     smc->default_caps.caps[SPAPR_CAP_HPT_MAXPAGESIZE] = 16; /* 64kiB */
     smc->default_caps.caps[SPAPR_CAP_NESTED_KVM_HV] = SPAPR_CAP_OFF;
     smc->default_caps.caps[SPAPR_CAP_LARGE_DECREMENTER] = SPAPR_CAP_ON;
-    smc->default_caps.caps[SPAPR_CAP_CCF_ASSIST] = SPAPR_CAP_OFF;
+    smc->default_caps.caps[SPAPR_CAP_CCF_ASSIST] = SPAPR_CAP_ON;
     spapr_caps_add_properties(smc, &error_abort);
     smc->irq = &spapr_irq_dual;
     smc->dr_phb_enabled = true;
@@ -4904,6 +4904,8 @@ static void spapr_machine_rhel810_class_options(MachineClass *mc)
                      hw_compat_rhel_8_1_len);
     compat_props_add(mc->compat_props, compat, G_N_ELEMENTS(compat));
 
+    /* from pseries-4.2 */
+    smc->default_caps.caps[SPAPR_CAP_CCF_ASSIST] = SPAPR_CAP_OFF;
 }
 
 DEFINE_SPAPR_MACHINE(rhel810, "rhel8.1.0", false);
diff --git a/hw/ppc/spapr_caps.c b/hw/ppc/spapr_caps.c
index 805f385..6e6fb28 100644
--- a/hw/ppc/spapr_caps.c
+++ b/hw/ppc/spapr_caps.c
@@ -492,11 +492,24 @@ static void cap_ccf_assist_apply(SpaprMachineState *spapr, uint8_t val,
     uint8_t kvm_val = kvmppc_get_cap_count_cache_flush_assist();
 
     if (tcg_enabled() && val) {
-        /* TODO - for now only allow broken for TCG */
-        error_setg(errp,
-"Requested count cache flush assist capability level not supported by tcg,"
-                   " try appending -machine cap-ccf-assist=off");
+        /* TCG doesn't implement anything here, but allow with a warning */
+        warn_report("TCG doesn't support requested feature, cap-ccf-assist=on");
     } else if (kvm_enabled() && (val > kvm_val)) {
+        uint8_t kvm_ibs = kvmppc_get_cap_safe_indirect_branch();
+
+        if (kvm_ibs == SPAPR_CAP_FIXED_CCD) {
+            /*
+             * If we don't have CCF assist on the host, the assist
+             * instruction is a harmless no-op.  It won't correctly
+             * implement the cache count flush *but* if we have
+             * count-cache-disabled in the host, that flush is
+             * unnnecessary.  So, specifically allow this case.  This
+             * allows us to have better performance on POWER9 DD2.3,
+             * while still working on POWER9 DD2.2 and POWER8 host
+             * cpus.
+             */
+            return;
+        }
         error_setg(errp,
 "Requested count cache flush assist capability level not supported by kvm,"
                    " try appending -machine cap-ccf-assist=off");
-- 
1.8.3.1