[PATCH v1 3/8] x86/svm: Remove lazy FPU support

[Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>]

Remove lazy FPU support from the SVM code since fully_eager_fpu is now
always true.

No functional change intended.

Signed-off-by: Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>
---
 xen/arch/x86/hvm/svm/nestedsvm.c         | 67 +-------------------
 xen/arch/x86/hvm/svm/svm.c               | 81 +-----------------------
 xen/arch/x86/hvm/svm/vmcb.c              |  4 +-
 xen/arch/x86/include/asm/hvm/svm-types.h |  6 --
 4 files changed, 4 insertions(+), 154 deletions(-)

diff --git a/xen/arch/x86/hvm/svm/nestedsvm.c b/xen/arch/x86/hvm/svm/nestedsvm.c
index a63ec613465f..ef6fa5d23b67 100644
--- a/xen/arch/x86/hvm/svm/nestedsvm.c
+++ b/xen/arch/x86/hvm/svm/nestedsvm.c
@@ -165,58 +165,6 @@ int cf_check nsvm_vcpu_reset(struct vcpu *v)
     return 0;
 }
 
-static uint64_t nestedsvm_fpu_vmentry(uint64_t n1cr0,
-    struct vmcb_struct *vvmcb,
-    struct vmcb_struct *n1vmcb, struct vmcb_struct *n2vmcb)
-{
-    uint64_t vcr0;
-
-    vcr0 = vvmcb->_cr0;
-    if ( !(n1cr0 & X86_CR0_TS) && (n1vmcb->_cr0 & X86_CR0_TS) )
-    {
-        /*
-         * svm_fpu_leave() run while l1 guest was running.
-         * Sync FPU state with l2 guest.
-         */
-        vcr0 |= X86_CR0_TS;
-        n2vmcb->_exception_intercepts |= (1U << X86_EXC_NM);
-    }
-    else if ( !(vcr0 & X86_CR0_TS) && (n2vmcb->_cr0 & X86_CR0_TS) )
-    {
-        /*
-         * svm_fpu_enter() run while l1 guest was running.
-         * Sync FPU state with l2 guest.
-         */
-        vcr0 &= ~X86_CR0_TS;
-        n2vmcb->_exception_intercepts &= ~(1U << X86_EXC_NM);
-    }
-
-    return vcr0;
-}
-
-static void nestedsvm_fpu_vmexit(struct vmcb_struct *n1vmcb,
-    struct vmcb_struct *n2vmcb, uint64_t n1cr0, uint64_t guest_cr0)
-{
-    if ( !(guest_cr0 & X86_CR0_TS) && (n2vmcb->_cr0 & X86_CR0_TS) )
-    {
-        /*
-         * svm_fpu_leave() run while l2 guest was running.
-         * Sync FPU state with l1 guest.
-         */
-        n1vmcb->_cr0 |= X86_CR0_TS;
-        n1vmcb->_exception_intercepts |= (1U << X86_EXC_NM);
-    }
-    else if ( !(n1cr0 & X86_CR0_TS) && (n1vmcb->_cr0 & X86_CR0_TS) )
-    {
-        /*
-         * svm_fpu_enter() run while l2 guest was running.
-         * Sync FPU state with l1 guest.
-         */
-        n1vmcb->_cr0 &= ~X86_CR0_TS;
-        n1vmcb->_exception_intercepts &= ~(1U << X86_EXC_NM);
-    }
-}
-
 static int nsvm_vcpu_hostsave(struct vcpu *v, unsigned int inst_len)
 {
     struct nestedsvm *svm = &vcpu_nestedsvm(v);
@@ -246,7 +194,6 @@ static int nsvm_vcpu_hostsave(struct vcpu *v, unsigned int 
inst_len)
 static int nsvm_vcpu_hostrestore(struct vcpu *v, struct cpu_user_regs *regs)
 {
     struct nestedvcpu *nv = &vcpu_nestedhvm(v);
-    struct nestedsvm *svm = &vcpu_nestedsvm(v);
     struct vmcb_struct *n1vmcb, *n2vmcb;
     int rc;
 
@@ -281,8 +228,6 @@ static int nsvm_vcpu_hostrestore(struct vcpu *v, struct 
cpu_user_regs *regs)
         gdprintk(XENLOG_ERR, "hvm_set_cr4 failed, rc: %u\n", rc);
 
     /* CR0 */
-    nestedsvm_fpu_vmexit(n1vmcb, n2vmcb,
-        svm->ns_cr0, v->arch.hvm.guest_cr[0]);
     v->arch.hvm.guest_cr[0] = n1vmcb->_cr0 | X86_CR0_PE;
     n1vmcb->rflags &= ~X86_EFLAGS_VM;
     rc = hvm_set_cr0(n1vmcb->_cr0 | X86_CR0_PE, true);
@@ -290,7 +235,6 @@ static int nsvm_vcpu_hostrestore(struct vcpu *v, struct 
cpu_user_regs *regs)
         hvm_inject_hw_exception(X86_EXC_GP, 0);
     if ( rc != X86EMUL_OKAY )
         gdprintk(XENLOG_ERR, "hvm_set_cr0 failed, rc: %u\n", rc);
-    svm->ns_cr0 = v->arch.hvm.guest_cr[0];
 
     /* CR2 */
     v->arch.hvm.guest_cr[2] = n1vmcb->_cr2;
@@ -418,7 +362,6 @@ static int nsvm_vmcb_prepare4vmrun(struct vcpu *v, struct 
cpu_user_regs *regs)
     struct vmcb_struct *ns_vmcb, *n1vmcb, *n2vmcb;
     vmcbcleanbits_t clean = {};
     int rc;
-    uint64_t cr0;
 
     ns_vmcb = nv->nv_vvmcx;
     n1vmcb = nv->nv_n1vmcx;
@@ -452,7 +395,6 @@ static int nsvm_vmcb_prepare4vmrun(struct vcpu *v, struct 
cpu_user_regs *regs)
      *   safed here.
      * The overhead comes from (ordered from highest to lowest):
      * - svm_ctxt_switch_to (CPU context switching)
-     * - svm_fpu_enter, svm_fpu_leave (lazy FPU switching)
      * - emulated CLGI (clears VINTR intercept)
      * - host clears VINTR intercept
      * Test results show that the overhead is high enough that the
@@ -551,10 +493,8 @@ static int nsvm_vmcb_prepare4vmrun(struct vcpu *v, struct 
cpu_user_regs *regs)
         gdprintk(XENLOG_ERR, "hvm_set_cr4 failed, rc: %u\n", rc);
 
     /* CR0 */
-    svm->ns_cr0 = v->arch.hvm.guest_cr[0];
-    cr0 = nestedsvm_fpu_vmentry(svm->ns_cr0, ns_vmcb, n1vmcb, n2vmcb);
     v->arch.hvm.guest_cr[0] = ns_vmcb->_cr0;
-    rc = hvm_set_cr0(cr0, true);
+    rc = hvm_set_cr0(ns_vmcb->_cr0, true);
     if ( rc == X86EMUL_EXCEPTION )
         hvm_inject_hw_exception(X86_EXC_GP, 0);
     if ( rc != X86EMUL_OKAY )
@@ -1305,11 +1245,6 @@ nestedsvm_check_intercepts(struct vcpu *v, struct 
cpu_user_regs *regs,
     case VMEXIT_INTR:
     case VMEXIT_NMI:
         return NESTEDHVM_VMEXIT_HOST;
-    case VMEXIT_EXCEPTION_NM:
-        /* Host must handle lazy fpu context switching first.
-         * Then inject the VMEXIT if L1 guest intercepts this.
-         */
-        return NESTEDHVM_VMEXIT_HOST;
 
     case VMEXIT_NPF:
         if ( nestedhvm_paging_mode_hap(v) )
diff --git a/xen/arch/x86/hvm/svm/svm.c b/xen/arch/x86/hvm/svm/svm.c
index 243c41fb13a8..2546705d245c 100644
--- a/xen/arch/x86/hvm/svm/svm.c
+++ b/xen/arch/x86/hvm/svm/svm.c
@@ -104,38 +104,6 @@ static void cf_check svm_cpu_down(void)
     write_efer(read_efer() & ~EFER_SVME);
 }
 
-static void svm_fpu_enter(struct vcpu *v)
-{
-    struct vmcb_struct *n1vmcb = vcpu_nestedhvm(v).nv_n1vmcx;
-
-    vcpu_restore_fpu_lazy(v);
-    vmcb_set_exception_intercepts(
-        n1vmcb,
-        vmcb_get_exception_intercepts(n1vmcb) & ~(1U << X86_EXC_NM));
-}
-
-static void cf_check svm_fpu_leave(struct vcpu *v)
-{
-    struct vmcb_struct *n1vmcb = vcpu_nestedhvm(v).nv_n1vmcx;
-
-    ASSERT(!v->fpu_dirtied);
-    ASSERT(read_cr0() & X86_CR0_TS);
-
-    /*
-     * If the guest does not have TS enabled then we must cause and handle an
-     * exception on first use of the FPU. If the guest *does* have TS enabled
-     * then this is not necessary: no FPU activity can occur until the guest
-     * clears CR0.TS, and we will initialise the FPU when that happens.
-     */
-    if ( !(v->arch.hvm.guest_cr[0] & X86_CR0_TS) )
-    {
-        vmcb_set_exception_intercepts(
-            n1vmcb,
-            vmcb_get_exception_intercepts(n1vmcb) | (1U << X86_EXC_NM));
-        vmcb_set_cr0(n1vmcb, vmcb_get_cr0(n1vmcb) | X86_CR0_TS);
-    }
-}
-
 static void cf_check svm_update_guest_cr(
     struct vcpu *v, unsigned int cr, unsigned int flags)
 {
@@ -145,20 +113,6 @@ static void cf_check svm_update_guest_cr(
     switch ( cr )
     {
     case 0:
-    {
-        unsigned long hw_cr0_mask = 0;
-
-        if ( !(v->arch.hvm.guest_cr[0] & X86_CR0_TS) )
-        {
-            if ( v != current )
-            {
-                if ( !v->arch.fully_eager_fpu )
-                    hw_cr0_mask |= X86_CR0_TS;
-            }
-            else if ( vmcb_get_cr0(vmcb) & X86_CR0_TS )
-                svm_fpu_enter(v);
-        }
-
         if ( paging_mode_hap(v->domain) )
         {
             uint32_t intercepts = vmcb_get_cr_intercepts(vmcb);
@@ -169,12 +123,12 @@ static void cf_check svm_update_guest_cr(
                vmcb_set_cr_intercepts(vmcb, intercepts | 
CR_INTERCEPT_CR3_WRITE);
         }
 
-        value = v->arch.hvm.guest_cr[0] | hw_cr0_mask;
+        value = v->arch.hvm.guest_cr[0];
         if ( paging_mode_shadow(v->domain) )
             value |= X86_CR0_PG | X86_CR0_WP;
         vmcb_set_cr0(vmcb, value);
         break;
-    }
+
     case 2:
         vmcb_set_cr2(vmcb, v->arch.hvm.guest_cr[2]);
         break;
@@ -909,9 +863,6 @@ static void cf_check svm_ctxt_switch_from(struct vcpu *v)
     if ( unlikely((read_efer() & EFER_SVME) == 0) )
         return;
 
-    if ( !v->arch.fully_eager_fpu )
-        svm_fpu_leave(v);
-
     svm_save_dr(v);
     svm_tsc_ratio_save(v);
 
@@ -1678,28 +1629,6 @@ static void svm_do_nested_pgfault(struct vcpu *v,
     domain_crash(v->domain);
 }
 
-static void cf_check svm_fpu_dirty_intercept(void)
-{
-    struct vcpu *v = current;
-    struct vmcb_struct *vmcb = v->arch.hvm.svm.vmcb;
-    struct vmcb_struct *n1vmcb = vcpu_nestedhvm(v).nv_n1vmcx;
-
-    svm_fpu_enter(v);
-
-    if ( vmcb != n1vmcb )
-    {
-       /* Check if l1 guest must make FPU ready for the l2 guest */
-       if ( v->arch.hvm.guest_cr[0] & X86_CR0_TS )
-           hvm_inject_hw_exception(X86_EXC_NM, X86_EVENT_NO_EC);
-       else
-           vmcb_set_cr0(n1vmcb, vmcb_get_cr0(n1vmcb) & ~X86_CR0_TS);
-       return;
-    }
-
-    if ( !(v->arch.hvm.guest_cr[0] & X86_CR0_TS) )
-        vmcb_set_cr0(vmcb, vmcb_get_cr0(vmcb) & ~X86_CR0_TS);
-}
-
 static void svm_vmexit_do_cr_access(
     struct vmcb_struct *vmcb, struct cpu_user_regs *regs)
 {
@@ -2459,7 +2388,6 @@ static struct hvm_function_table __initdata_cf_clobber 
svm_function_table = {
     .update_guest_cr      = svm_update_guest_cr,
     .update_guest_efer    = svm_update_guest_efer,
     .cpuid_policy_changed = svm_cpuid_policy_changed,
-    .fpu_leave            = svm_fpu_leave,
     .set_guest_pat        = svm_set_guest_pat,
     .get_guest_pat        = svm_get_guest_pat,
     .set_tsc_offset       = svm_set_tsc_offset,
@@ -2469,7 +2397,6 @@ static struct hvm_function_table __initdata_cf_clobber 
svm_function_table = {
     .get_pending_event    = svm_get_pending_event,
     .invlpg               = svm_invlpg,
     .wbinvd_intercept     = svm_wbinvd_intercept,
-    .fpu_dirty_intercept  = svm_fpu_dirty_intercept,
     .msr_read_intercept   = svm_msr_read_intercept,
     .msr_write_intercept  = svm_msr_write_intercept,
 #ifdef CONFIG_VM_EVENT
@@ -2783,10 +2710,6 @@ void asmlinkage svm_vmexit_handler(void)
         }
         break;
 
-    case VMEXIT_EXCEPTION_NM:
-        svm_fpu_dirty_intercept();
-        break;
-
     case VMEXIT_EXCEPTION_PF:
     {
         unsigned long va = vmcb->ei.exc.cr2;
diff --git a/xen/arch/x86/hvm/svm/vmcb.c b/xen/arch/x86/hvm/svm/vmcb.c
index e583ef8548c7..5ed7123d9a69 100644
--- a/xen/arch/x86/hvm/svm/vmcb.c
+++ b/xen/arch/x86/hvm/svm/vmcb.c
@@ -138,9 +138,7 @@ static int construct_vmcb(struct vcpu *v)
 
     paging_update_paging_modes(v);
 
-    vmcb->_exception_intercepts =
-        HVM_TRAP_MASK |
-        (v->arch.fully_eager_fpu ? 0 : (1U << X86_EXC_NM));
+    vmcb->_exception_intercepts = HVM_TRAP_MASK;
 
     if ( paging_mode_hap(v->domain) )
     {
diff --git a/xen/arch/x86/include/asm/hvm/svm-types.h 
b/xen/arch/x86/include/asm/hvm/svm-types.h
index 051b235d8f69..3ede62cade80 100644
--- a/xen/arch/x86/include/asm/hvm/svm-types.h
+++ b/xen/arch/x86/include/asm/hvm/svm-types.h
@@ -65,12 +65,6 @@ struct nestedsvm {
     /* Shadow io permission map */
     unsigned long *ns_iomap;
 
-     /*
-      * Cached guest_cr[0] of l1 guest while l2 guest runs.  Needed to handle
-      * FPU context switching.
-      */
-    uint64_t ns_cr0;
-
     /*
      * Cache guest cr3/host cr3 the guest sets up for the l2 guest.
      * Used by Shadow-on-Shadow and Nested-on-Nested.
-- 
2.53.0