Patchwork [v2,uq/master] kvm: x86: Save/restore FPU OP, IP and DP

login
register
mail settings
Submitter Jan Kiszka
Date June 15, 2011, 1:17 p.m.
Message ID <4DF8B0E6.1080806@siemens.com>
Download mbox | patch
Permalink /patch/100516/
State New
Headers show

Comments

Jan Kiszka - June 15, 2011, 1:17 p.m.
These FPU states are properly maintained by KVM but not yet by TCG. So
far we unconditionally set them to 0 in the guest which may cause
state corruptions, though not with modern guests.

To avoid breaking backward migration, use a conditional subsection that
is only written if any of the three fields is non-zero. The guest's
FNINIT clears them frequently, and cleared IA32_MISC_ENABLE MSR[2]
reduces the probability of non-zero values further so that this
subsection is not expected to restrict migration in any common scenario.

Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
---

Changes in v2:
 - use conditional subsection

 target-i386/cpu.h     |    4 ++++
 target-i386/kvm.c     |   20 +++++++++++++++-----
 target-i386/machine.c |   23 +++++++++++++++++++++++
 3 files changed, 42 insertions(+), 5 deletions(-)
Christophe Fergeau - June 16, 2011, 9:42 a.m.
On Wed, Jun 15, 2011 at 03:17:26PM +0200, Jan Kiszka wrote:
> These FPU states are properly maintained by KVM but not yet by TCG. So
> far we unconditionally set them to 0 in the guest which may cause
> state corruptions, though not with modern guests.

And this one's fine too from a compilation point of view on f15/gcc4.6

Christophe
Avi Kivity - June 19, 2011, 12:58 p.m.
On 06/15/2011 04:17 PM, Jan Kiszka wrote:
> These FPU states are properly maintained by KVM but not yet by TCG. So
> far we unconditionally set them to 0 in the guest which may cause
> state corruptions, though not with modern guests.
>
> To avoid breaking backward migration, use a conditional subsection that
> is only written if any of the three fields is non-zero. The guest's
> FNINIT clears them frequently, and cleared IA32_MISC_ENABLE MSR[2]
> reduces the probability of non-zero values further so that this
> subsection is not expected to restrict migration in any common scenario.

Thanks, applied to uq/master.  A test migration showed that the 
subsection in fact did not show up.

Patch

diff --git a/target-i386/cpu.h b/target-i386/cpu.h
index 9c3340d..cdf68ff 100644
--- a/target-i386/cpu.h
+++ b/target-i386/cpu.h
@@ -641,6 +641,10 @@  typedef struct CPUX86State {
     uint16_t fpuc;
     uint8_t fptags[8];   /* 0 = valid, 1 = empty */
     FPReg fpregs[8];
+    /* KVM-only so far */
+    uint16_t fpop;
+    uint64_t fpip;
+    uint64_t fpdp;
 
     /* emulator internal variables */
     float_status fp_status;
diff --git a/target-i386/kvm.c b/target-i386/kvm.c
index 5ebb054..938e0a3 100644
--- a/target-i386/kvm.c
+++ b/target-i386/kvm.c
@@ -718,6 +718,9 @@  static int kvm_put_fpu(CPUState *env)
     fpu.fsw = env->fpus & ~(7 << 11);
     fpu.fsw |= (env->fpstt & 7) << 11;
     fpu.fcw = env->fpuc;
+    fpu.last_opcode = env->fpop;
+    fpu.last_ip = env->fpip;
+    fpu.last_dp = env->fpdp;
     for (i = 0; i < 8; ++i) {
         fpu.ftwx |= (!env->fptags[i]) << i;
     }
@@ -740,7 +743,7 @@  static int kvm_put_xsave(CPUState *env)
 {
     int i, r;
     struct kvm_xsave* xsave;
-    uint16_t cwd, swd, twd, fop;
+    uint16_t cwd, swd, twd;
 
     if (!kvm_has_xsave()) {
         return kvm_put_fpu(env);
@@ -748,7 +751,7 @@  static int kvm_put_xsave(CPUState *env)
 
     xsave = qemu_memalign(4096, sizeof(struct kvm_xsave));
     memset(xsave, 0, sizeof(struct kvm_xsave));
-    cwd = swd = twd = fop = 0;
+    cwd = swd = twd = 0;
     swd = env->fpus & ~(7 << 11);
     swd |= (env->fpstt & 7) << 11;
     cwd = env->fpuc;
@@ -756,7 +759,9 @@  static int kvm_put_xsave(CPUState *env)
         twd |= (!env->fptags[i]) << i;
     }
     xsave->region[0] = (uint32_t)(swd << 16) + cwd;
-    xsave->region[1] = (uint32_t)(fop << 16) + twd;
+    xsave->region[1] = (uint32_t)(env->fpop << 16) + twd;
+    memcpy(&xsave->region[XSAVE_CWD_RIP], &env->fpip, sizeof(env->fpip));
+    memcpy(&xsave->region[XSAVE_CWD_RDP], &env->fpdp, sizeof(env->fpdp));
     memcpy(&xsave->region[XSAVE_ST_SPACE], env->fpregs,
             sizeof env->fpregs);
     memcpy(&xsave->region[XSAVE_XMM_SPACE], env->xmm_regs,
@@ -921,6 +926,9 @@  static int kvm_get_fpu(CPUState *env)
     env->fpstt = (fpu.fsw >> 11) & 7;
     env->fpus = fpu.fsw;
     env->fpuc = fpu.fcw;
+    env->fpop = fpu.last_opcode;
+    env->fpip = fpu.last_ip;
+    env->fpdp = fpu.last_dp;
     for (i = 0; i < 8; ++i) {
         env->fptags[i] = !((fpu.ftwx >> i) & 1);
     }
@@ -935,7 +943,7 @@  static int kvm_get_xsave(CPUState *env)
 {
     struct kvm_xsave* xsave;
     int ret, i;
-    uint16_t cwd, swd, twd, fop;
+    uint16_t cwd, swd, twd;
 
     if (!kvm_has_xsave()) {
         return kvm_get_fpu(env);
@@ -951,13 +959,15 @@  static int kvm_get_xsave(CPUState *env)
     cwd = (uint16_t)xsave->region[0];
     swd = (uint16_t)(xsave->region[0] >> 16);
     twd = (uint16_t)xsave->region[1];
-    fop = (uint16_t)(xsave->region[1] >> 16);
+    env->fpop = (uint16_t)(xsave->region[1] >> 16);
     env->fpstt = (swd >> 11) & 7;
     env->fpus = swd;
     env->fpuc = cwd;
     for (i = 0; i < 8; ++i) {
         env->fptags[i] = !((twd >> i) & 1);
     }
+    memcpy(&env->fpip, &xsave->region[XSAVE_CWD_RIP], sizeof(env->fpip));
+    memcpy(&env->fpdp, &xsave->region[XSAVE_CWD_RDP], sizeof(env->fpdp));
     env->mxcsr = xsave->region[XSAVE_MXCSR];
     memcpy(env->fpregs, &xsave->region[XSAVE_ST_SPACE],
             sizeof env->fpregs);
diff --git a/target-i386/machine.c b/target-i386/machine.c
index bbeae88..d22a731 100644
--- a/target-i386/machine.c
+++ b/target-i386/machine.c
@@ -290,6 +290,26 @@  static const VMStateDescription vmstate_async_pf_msr = {
     }
 };
 
+static bool fpop_ip_dp_needed(void *opaque)
+{
+    CPUState *env = opaque;
+
+    return env->fpop != 0 || env->fpip != 0 || env->fpdp != 0;
+}
+
+static const VMStateDescription vmstate_fpop_ip_dp = {
+    .name = "cpu/fpop_ip_dp",
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .minimum_version_id_old = 1,
+    .fields      = (VMStateField []) {
+        VMSTATE_UINT16_V(fpop, CPUState, 13),
+        VMSTATE_UINT64_V(fpip, CPUState, 13),
+        VMSTATE_UINT64_V(fpdp, CPUState, 13),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
 static const VMStateDescription vmstate_cpu = {
     .name = "cpu",
     .version_id = CPU_SAVE_VERSION,
@@ -398,6 +418,9 @@  static const VMStateDescription vmstate_cpu = {
             .vmsd = &vmstate_async_pf_msr,
             .needed = async_pf_msr_needed,
         } , {
+            .vmsd = &vmstate_fpop_ip_dp,
+            .needed = fpop_ip_dp_needed,
+        } , {
             /* empty */
         }
     }