]> err.no Git - linux-2.6/commitdiff
KVM: Make kvm host use the paravirt clocksource structs
authorGerd Hoffmann <kraxel@redhat.com>
Tue, 3 Jun 2008 14:17:31 +0000 (16:17 +0200)
committerAvi Kivity <avi@qumranet.com>
Tue, 24 Jun 2008 18:02:32 +0000 (21:02 +0300)
This patch updates the kvm host code to use the pvclock structs.
It also makes the paravirt clock compatible with Xen.

Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
arch/x86/kvm/x86.c
include/asm-x86/kvm_host.h

index b08812d6b34c8312ebdd8f2283c44679002d7e79..63a77caa59f11bf55a8655b857aa9c541e4168a6 100644 (file)
@@ -492,8 +492,8 @@ static int do_set_msr(struct kvm_vcpu *vcpu, unsigned index, u64 *data)
 static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock)
 {
        static int version;
-       struct kvm_wall_clock wc;
-       struct timespec wc_ts;
+       struct pvclock_wall_clock wc;
+       struct timespec now, sys, boot;
 
        if (!wall_clock)
                return;
@@ -502,10 +502,19 @@ static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock)
 
        kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
 
-       wc_ts = current_kernel_time();
-       wc.wc_sec = wc_ts.tv_sec;
-       wc.wc_nsec = wc_ts.tv_nsec;
-       wc.wc_version = version;
+       /*
+        * The guest calculates current wall clock time by adding
+        * system time (updated by kvm_write_guest_time below) to the
+        * wall clock specified here.  guest system time equals host
+        * system time for us, thus we must fill in host boot time here.
+        */
+       now = current_kernel_time();
+       ktime_get_ts(&sys);
+       boot = ns_to_timespec(timespec_to_ns(&now) - timespec_to_ns(&sys));
+
+       wc.sec = boot.tv_sec;
+       wc.nsec = boot.tv_nsec;
+       wc.version = version;
 
        kvm_write_guest(kvm, wall_clock, &wc, sizeof(wc));
 
@@ -513,6 +522,45 @@ static void kvm_write_wall_clock(struct kvm *kvm, gpa_t wall_clock)
        kvm_write_guest(kvm, wall_clock, &version, sizeof(version));
 }
 
+static uint32_t div_frac(uint32_t dividend, uint32_t divisor)
+{
+       uint32_t quotient, remainder;
+
+       /* Don't try to replace with do_div(), this one calculates
+        * "(dividend << 32) / divisor" */
+       __asm__ ( "divl %4"
+                 : "=a" (quotient), "=d" (remainder)
+                 : "0" (0), "1" (dividend), "r" (divisor) );
+       return quotient;
+}
+
+static void kvm_set_time_scale(uint32_t tsc_khz, struct pvclock_vcpu_time_info *hv_clock)
+{
+       uint64_t nsecs = 1000000000LL;
+       int32_t  shift = 0;
+       uint64_t tps64;
+       uint32_t tps32;
+
+       tps64 = tsc_khz * 1000LL;
+       while (tps64 > nsecs*2) {
+               tps64 >>= 1;
+               shift--;
+       }
+
+       tps32 = (uint32_t)tps64;
+       while (tps32 <= (uint32_t)nsecs) {
+               tps32 <<= 1;
+               shift++;
+       }
+
+       hv_clock->tsc_shift = shift;
+       hv_clock->tsc_to_system_mul = div_frac(nsecs, tps32);
+
+       pr_debug("%s: tsc_khz %u, tsc_shift %d, tsc_mul %u\n",
+                __FUNCTION__, tsc_khz, hv_clock->tsc_shift,
+                hv_clock->tsc_to_system_mul);
+}
+
 static void kvm_write_guest_time(struct kvm_vcpu *v)
 {
        struct timespec ts;
@@ -523,6 +571,11 @@ static void kvm_write_guest_time(struct kvm_vcpu *v)
        if ((!vcpu->time_page))
                return;
 
+       if (unlikely(vcpu->hv_clock_tsc_khz != tsc_khz)) {
+               kvm_set_time_scale(tsc_khz, &vcpu->hv_clock);
+               vcpu->hv_clock_tsc_khz = tsc_khz;
+       }
+
        /* Keep irq disabled to prevent changes to the clock */
        local_irq_save(flags);
        kvm_get_msr(v, MSR_IA32_TIME_STAMP_COUNTER,
@@ -537,14 +590,14 @@ static void kvm_write_guest_time(struct kvm_vcpu *v)
        /*
         * The interface expects us to write an even number signaling that the
         * update is finished. Since the guest won't see the intermediate
-        * state, we just write "2" at the end
+        * state, we just increase by 2 at the end.
         */
-       vcpu->hv_clock.version = 2;
+       vcpu->hv_clock.version += 2;
 
        shared_kaddr = kmap_atomic(vcpu->time_page, KM_USER0);
 
        memcpy(shared_kaddr + vcpu->time_offset, &vcpu->hv_clock,
-               sizeof(vcpu->hv_clock));
+              sizeof(vcpu->hv_clock));
 
        kunmap_atomic(shared_kaddr, KM_USER0);
 
@@ -599,10 +652,6 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 data)
                /* ...but clean it before doing the actual write */
                vcpu->arch.time_offset = data & ~(PAGE_MASK | 1);
 
-               vcpu->arch.hv_clock.tsc_to_system_mul =
-                                       clocksource_khz2mult(tsc_khz, 22);
-               vcpu->arch.hv_clock.tsc_shift = 22;
-
                down_read(&current->mm->mmap_sem);
                vcpu->arch.time_page =
                                gfn_to_page(vcpu->kvm, data >> PAGE_SHIFT);
index 1d8cd01fa5147221ccf6e598b8d8f0d3a100bf3e..844f2a89afbcdb28a8a6272fd9826b8a24777351 100644 (file)
@@ -18,6 +18,7 @@
 #include <linux/kvm_para.h>
 #include <linux/kvm_types.h>
 
+#include <asm/pvclock-abi.h>
 #include <asm/desc.h>
 
 #define KVM_MAX_VCPUS 16
@@ -282,7 +283,8 @@ struct kvm_vcpu_arch {
        struct x86_emulate_ctxt emulate_ctxt;
 
        gpa_t time;
-       struct kvm_vcpu_time_info hv_clock;
+       struct pvclock_vcpu_time_info hv_clock;
+       unsigned int hv_clock_tsc_khz;
        unsigned int time_offset;
        struct page *time_page;
 };