x86: Workaorund for the KVM GS_BASE MSR save/restore issue.

In some versions of the kvm module in Ubuntu, the host's GS_BASE MSR is not save/restored correctly
when running guests on 64-bit hosts if the qemu/emulator is 32-bit.
This patch implements a workaround in the emulator code.

Change-Id: If3ebe3cb49f377c5d0547c75f6ac3a3ceacdc375
Signed-off-by: Yunhong Jiang <yunhong.jiang@intel.com>
Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>

3 files changed, 177 insertions, 0 deletions

diff --git a/target-i386/kvm-gs-restore.c b/target-i386/kvm-gs-restore.c
new file mode 100644
index 0000000..a0ff565
--- /dev/null
+++ b/target-i386/kvm-gs-restore.c
@@ -0,0 +1,144 @@
+#include <sys/types.h>
+#include <sys/ioctl.h>
+#include "qemu-common.h"
+
+#ifdef CONFIG_KVM_GS_RESTORE
+
+#define INVALID_GS_REG  0xffff
+#define KVM_GS_RESTORE_NODETECTED 0x1
+#define KVM_GS_RESTORE_NO 0x2
+#define KVM_GS_RESTORE_YES 0x3
+int initial_gs = INVALID_GS_REG;
+int gs_need_restore = KVM_GS_RESTORE_NODETECTED;
+
+static void restoregs(int gs)
+{
+    asm("movl %0, %%gs"::"r"(gs));
+}
+
+static unsigned int _getgs()
+{
+    unsigned int gs = 0;
+    asm("movl %%gs,%0" :"=r"(gs):);
+    return gs;
+}
+
+/* No fprintf or any system call before the gs is restored successfully */
+static void check_and_restore_gs(void)
+{
+    if (gs_need_restore == KVM_GS_RESTORE_NO)
+        return;
+
+    restoregs(initial_gs);
+}
+
+static struct sigaction old_alarm_act, old_gio_act, old_pipe_act,old_usr1_act, old_chld_act;
+static void temp_sig_handler(int host_signum)
+{
+    /* !!! must restore gs firstly */
+    check_and_restore_gs();
+    switch (host_signum)
+    {
+    case SIGALRM:
+        old_alarm_act.sa_handler(host_signum);
+        break;
+
+    case SIGIO:
+        old_gio_act.sa_handler(host_signum);
+        break;
+
+    case SIGUSR1:
+        old_usr1_act.sa_handler(host_signum);
+        break;
+
+    case SIGPIPE:
+        old_pipe_act.sa_handler(host_signum);
+        break;
+
+    case SIGCHLD:
+        old_chld_act.sa_handler(host_signum);
+        break;
+
+    default:
+        fprintf(stderr, "Not take signal %x!!\n", host_signum);
+        break;
+    }
+}
+
+static int sig_taken = 0;
+static int take_signal_handler(void)
+{
+    struct sigaction act;
+    int ret;
+
+    if (gs_need_restore == KVM_GS_RESTORE_NO)
+        return 0;
+    if (sig_taken)
+        return 0;
+
+    sig_taken = 1;
+    sigfillset(&act.sa_mask);
+    act.sa_flags = 0;
+    act.sa_handler = temp_sig_handler;
+    /* Did we missed any other signal ? */
+    sigaction(SIGALRM, &act, &old_alarm_act);
+    sigaction(SIGIO, &act, &old_gio_act);
+    sigaction(SIGUSR1, &act, &old_usr1_act);
+    sigaction(SIGPIPE, &act, &old_pipe_act);
+    act.sa_flags = SA_NOCLDSTOP;
+    sigaction(SIGCHLD, &act, &old_chld_act);
+    return 1;
+}
+
+int gs_base_pre_run(void)
+{
+    if (unlikely(initial_gs == INVALID_GS_REG) )
+    {
+        initial_gs = _getgs();
+        /*
+         * As 2.6.35-28 lucid will get correct gs but clobbered GS_BASE
+         * we have to always re-write the gs base
+         */
+        if (initial_gs == 0x0)
+            gs_need_restore = KVM_GS_RESTORE_NO;
+        else
+            gs_need_restore = KVM_GS_RESTORE_YES;
+    }
+
+    take_signal_handler();
+    return 0;
+}
+
+int gs_base_post_run(void)
+{
+    check_and_restore_gs();
+    return 0;
+}
+
+/*
+ * ioctl may update errno, which is in thread local storage and
+ * requires gs register, we have to provide our own ioctl
+ * XXX should "call %%gs:$0x10" be replaced with call to vsyscall
+ * page, which is more generic and clean?
+ */
+int no_gs_ioctl(int fd, int type, void *arg)
+{
+    int ret=0;
+
+    asm(
+      "movl %3, %%edx;\n"
+      "movl %2, %%ecx;\n"
+      "movl %1, %%ebx;\n"
+      "movl $0x36, %%eax;\n"
+      "call *%%gs:0x10;\n"
+      "movl %%eax, %0\n"
+      : "=m"(ret)
+      :"m"(fd),"m"(type),"m"(arg)
+      :"%edx","%ecx","%eax","%ebx"
+      );
+
+    return ret;
+}
+
+#endif
+
diff --git a/target-i386/kvm-gs-restore.h b/target-i386/kvm-gs-restore.h
new file mode 100644
index 0000000..0947b33
--- /dev/null
+++ b/target-i386/kvm-gs-restore.h
@@ -0,0 +1,12 @@
+#ifndef _KVM_GS_RESTORE_H
+#define _KVM_GS_RESTORE_H
+
+#ifdef CONFIG_KVM_GS_RESTORE
+int no_gs_ioctl(int fd, int type, void *arg);
+int gs_base_post_run(void);
+int gs_base_pre_run(void);
+extern int gs_need_restore;
+#define KVM_GS_RESTORE_NO 0x2
+#endif
+
+#endif
diff --git a/target-i386/kvm.c b/target-i386/kvm.c
index 4b135b1..07efac0 100644
--- a/target-i386/kvm.c
+++ b/target-i386/kvm.c
@@ -26,6 +26,10 @@
 #include "cpu.h"
 #include "gdbstub.h"
 
+#ifdef CONFIG_KVM_GS_RESTORE
+#include "kvm-gs-restore.h"
+#endif
+
 //#define DEBUG_KVM
 
 #ifdef DEBUG_KVM
@@ -690,6 +694,16 @@ int kvm_arch_get_registers(CPUState *env)
     return 0;
 }
 
+int kvm_arch_vcpu_run(CPUState *env)
+{
+#ifdef CONFIG_KVM_GS_RESTORE
+    if (gs_need_restore  != KVM_GS_RESTORE_NO)
+        return no_gs_ioctl(env->kvm_fd, KVM_RUN, 0);
+    else
+#endif
+        return kvm_vcpu_ioctl(env, KVM_RUN, 0);
+}
+
 int kvm_arch_pre_run(CPUState *env, struct kvm_run *run)
 {
     /* Try to inject an interrupt if the guest can accept it */
@@ -721,11 +735,18 @@ int kvm_arch_pre_run(CPUState *env, struct kvm_run *run)
     dprintf("setting tpr\n");
     run->cr8 = cpu_get_apic_tpr(env);
 
+#ifdef CONFIG_KVM_GS_RESTORE
+    gs_base_pre_run();
+#endif
+
     return 0;
 }
 
 int kvm_arch_post_run(CPUState *env, struct kvm_run *run)
 {
+#ifdef CONFIG_KVM_GS_RESTORE
+    gs_base_post_run();
+#endif
     if (run->if_flag)
         env->eflags |= IF_MASK;
     else