x86: KVM guest: paravirtualized clocksource
authorGlauber de Oliveira Costa <gcosta@redhat.com>
Fri, 15 Feb 2008 19:52:48 +0000 (17:52 -0200)
committerAvi Kivity <avi@qumranet.com>
Sun, 27 Apr 2008 08:53:22 +0000 (11:53 +0300)
This is the guest part of kvm clock implementation
It does not do tsc-only timing, as tsc can have deltas
between cpus, and it did not seem worthy to me to keep
adjusting them.

We do use it, however, for fine-grained adjustment.

Other than that, time comes from the host.

[randy dunlap: add missing include]
[randy dunlap: disallow on Voyager or Visual WS]

Signed-off-by: Glauber de Oliveira Costa <gcosta@redhat.com>
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
arch/x86/Kconfig
arch/x86/kernel/Makefile
arch/x86/kernel/kvmclock.c [new file with mode: 0644]
arch/x86/kernel/setup_32.c
arch/x86/kernel/setup_64.c

index 2fadf794483d1f8dad597c34de6dafb7082b4383..40cedc255eda8c096529e0f649dc17e874edf970 100644 (file)
@@ -373,6 +373,17 @@ config VMI
          at the moment), by linking the kernel to a GPL-ed ROM module
          provided by the hypervisor.
 
+config KVM_CLOCK
+       bool "KVM paravirtualized clock"
+       select PARAVIRT
+       depends on !(X86_VISWS || X86_VOYAGER)
+       help
+         Turning on this option will allow you to run a paravirtualized clock
+         when running over the KVM hypervisor. Instead of relying on a PIT
+         (or probably other) emulation by the underlying device model, the host
+         provides the guest with timing infrastructure such as time of day, and
+         system time
+
 source "arch/x86/lguest/Kconfig"
 
 config PARAVIRT
index 90e092d0af0c639211932fc44a0f5dd959ec590b..483047a33024578c7f0d5b2c6bc511ca49955d2c 100644 (file)
@@ -80,6 +80,7 @@ obj-$(CONFIG_DEBUG_RODATA_TEST)       += test_rodata.o
 obj-$(CONFIG_DEBUG_NX_TEST)    += test_nx.o
 
 obj-$(CONFIG_VMI)              += vmi_32.o vmiclock_32.o
+obj-$(CONFIG_KVM_CLOCK)                += kvmclock.o
 obj-$(CONFIG_PARAVIRT)         += paravirt.o paravirt_patch_$(BITS).o
 
 ifdef CONFIG_INPUT_PCSPKR
diff --git a/arch/x86/kernel/kvmclock.c b/arch/x86/kernel/kvmclock.c
new file mode 100644 (file)
index 0000000..b999f5e
--- /dev/null
@@ -0,0 +1,160 @@
+/*  KVM paravirtual clock driver. A clocksource implementation
+    Copyright (C) 2008 Glauber de Oliveira Costa, Red Hat Inc.
+
+    This program is free software; you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation; either version 2 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program; if not, write to the Free Software
+    Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+*/
+
+#include <linux/clocksource.h>
+#include <linux/kvm_para.h>
+#include <asm/arch_hooks.h>
+#include <asm/msr.h>
+#include <asm/apic.h>
+#include <linux/percpu.h>
+
+#define KVM_SCALE 22
+
+static int kvmclock = 1;
+
+static int parse_no_kvmclock(char *arg)
+{
+       kvmclock = 0;
+       return 0;
+}
+early_param("no-kvmclock", parse_no_kvmclock);
+
+/* The hypervisor will put information about time periodically here */
+static DEFINE_PER_CPU_SHARED_ALIGNED(struct kvm_vcpu_time_info, hv_clock);
+#define get_clock(cpu, field) per_cpu(hv_clock, cpu).field
+
+static inline u64 kvm_get_delta(u64 last_tsc)
+{
+       int cpu = smp_processor_id();
+       u64 delta = native_read_tsc() - last_tsc;
+       return (delta * get_clock(cpu, tsc_to_system_mul)) >> KVM_SCALE;
+}
+
+static struct kvm_wall_clock wall_clock;
+static cycle_t kvm_clock_read(void);
+/*
+ * The wallclock is the time of day when we booted. Since then, some time may
+ * have elapsed since the hypervisor wrote the data. So we try to account for
+ * that with system time
+ */
+unsigned long kvm_get_wallclock(void)
+{
+       u32 wc_sec, wc_nsec;
+       u64 delta;
+       struct timespec ts;
+       int version, nsec;
+       int low, high;
+
+       low = (int)__pa(&wall_clock);
+       high = ((u64)__pa(&wall_clock) >> 32);
+
+       delta = kvm_clock_read();
+
+       native_write_msr(MSR_KVM_WALL_CLOCK, low, high);
+       do {
+               version = wall_clock.wc_version;
+               rmb();
+               wc_sec = wall_clock.wc_sec;
+               wc_nsec = wall_clock.wc_nsec;
+               rmb();
+       } while ((wall_clock.wc_version != version) || (version & 1));
+
+       delta = kvm_clock_read() - delta;
+       delta += wc_nsec;
+       nsec = do_div(delta, NSEC_PER_SEC);
+       set_normalized_timespec(&ts, wc_sec + delta, nsec);
+       /*
+        * Of all mechanisms of time adjustment I've tested, this one
+        * was the champion!
+        */
+       return ts.tv_sec + 1;
+}
+
+int kvm_set_wallclock(unsigned long now)
+{
+       return 0;
+}
+
+/*
+ * This is our read_clock function. The host puts an tsc timestamp each time
+ * it updates a new time. Without the tsc adjustment, we can have a situation
+ * in which a vcpu starts to run earlier (smaller system_time), but probes
+ * time later (compared to another vcpu), leading to backwards time
+ */
+static cycle_t kvm_clock_read(void)
+{
+       u64 last_tsc, now;
+       int cpu;
+
+       preempt_disable();
+       cpu = smp_processor_id();
+
+       last_tsc = get_clock(cpu, tsc_timestamp);
+       now = get_clock(cpu, system_time);
+
+       now += kvm_get_delta(last_tsc);
+       preempt_enable();
+
+       return now;
+}
+static struct clocksource kvm_clock = {
+       .name = "kvm-clock",
+       .read = kvm_clock_read,
+       .rating = 400,
+       .mask = CLOCKSOURCE_MASK(64),
+       .mult = 1 << KVM_SCALE,
+       .shift = KVM_SCALE,
+       .flags = CLOCK_SOURCE_IS_CONTINUOUS,
+};
+
+static int kvm_register_clock(void)
+{
+       int cpu = smp_processor_id();
+       int low, high;
+       low = (int)__pa(&per_cpu(hv_clock, cpu)) | 1;
+       high = ((u64)__pa(&per_cpu(hv_clock, cpu)) >> 32);
+
+       return native_write_msr_safe(MSR_KVM_SYSTEM_TIME, low, high);
+}
+
+static void kvm_setup_secondary_clock(void)
+{
+       /*
+        * Now that the first cpu already had this clocksource initialized,
+        * we shouldn't fail.
+        */
+       WARN_ON(kvm_register_clock());
+       /* ok, done with our trickery, call native */
+       setup_secondary_APIC_clock();
+}
+
+void __init kvmclock_init(void)
+{
+       if (!kvm_para_available())
+               return;
+
+       if (kvmclock && kvm_para_has_feature(KVM_FEATURE_CLOCKSOURCE)) {
+               if (kvm_register_clock())
+                       return;
+               pv_time_ops.get_wallclock = kvm_get_wallclock;
+               pv_time_ops.set_wallclock = kvm_set_wallclock;
+               pv_time_ops.sched_clock = kvm_clock_read;
+               pv_apic_ops.setup_secondary_clock = kvm_setup_secondary_clock;
+               clocksource_register(&kvm_clock);
+       }
+}
index 44cc9b933932f2502cd2900fc4e3893513b1ea62..5a849ddd09ee95c409d217a9d05813d4d01efe44 100644 (file)
@@ -47,6 +47,7 @@
 #include <linux/pfn.h>
 #include <linux/pci.h>
 #include <linux/init_ohci1394_dma.h>
+#include <linux/kvm_para.h>
 
 #include <video/edid.h>
 
@@ -820,6 +821,10 @@ void __init setup_arch(char **cmdline_p)
 
        max_low_pfn = setup_memory();
 
+#ifdef CONFIG_KVM_CLOCK
+       kvmclock_init();
+#endif
+
 #ifdef CONFIG_VMI
        /*
         * Must be after max_low_pfn is determined, and before kernel
index 60e64c8eee92a503d3b88a96c142a3c7d1a950b0..8a9213c023effbabacda48c076650fec1cf38684 100644 (file)
@@ -42,6 +42,7 @@
 #include <linux/ctype.h>
 #include <linux/uaccess.h>
 #include <linux/init_ohci1394_dma.h>
+#include <linux/kvm_para.h>
 
 #include <asm/mtrr.h>
 #include <asm/uaccess.h>
@@ -384,6 +385,10 @@ void __init setup_arch(char **cmdline_p)
 
        io_delay_init();
 
+#ifdef CONFIG_KVM_CLOCK
+       kvmclock_init();
+#endif
+
 #ifdef CONFIG_SMP
        /* setup to use the early static init tables during kernel startup */
        x86_cpu_to_apicid_early_ptr = (void *)x86_cpu_to_apicid_init;