/* interruptibility state, as a result of execution of STI or MOV SS */
        int interruptibility;
 
-       bool restart; /* restart string instruction after writeback */
        bool perm_ok; /* do not check permissions if true */
 
        int exception; /* exception that happens during emulation or -1 */
 #endif
 
 int x86_decode_insn(struct x86_emulate_ctxt *ctxt);
+#define EMULATION_FAILED -1
+#define EMULATION_OK 0
+#define EMULATION_RESTART 1
 int x86_emulate_insn(struct x86_emulate_ctxt *ctxt);
 int emulator_task_switch(struct x86_emulate_ctxt *ctxt,
                         u16 tss_selector, int reason,
 
        ctxt->exception = vec;
        ctxt->error_code = error;
        ctxt->error_code_valid = valid;
-       ctxt->restart = false;
 }
 
 static void emulate_gp(struct x86_emulate_ctxt *ctxt, int err)
        struct opcode opcode, *g_mod012, *g_mod3;
        struct operand memop = { .type = OP_NONE };
 
-       /* we cannot decode insn before we complete previous rep insn */
-       WARN_ON(ctxt->restart);
-
        c->eip = ctxt->eip;
        c->fetch.start = c->fetch.end = c->eip;
        ctxt->cs_base = seg_base(ctxt, ops, VCPU_SREG_CS);
        }
 
        if (c->rep_prefix && (c->d & String)) {
-               ctxt->restart = true;
                /* All REP prefixes have the same first termination condition */
                if (address_mask(c, c->regs[VCPU_REGS_RCX]) == 0) {
-                       ctxt->restart = false;
                        ctxt->eip = c->eip;
                        goto done;
                }
                struct read_cache *r = &ctxt->decode.io_read;
                register_address_increment(c, &c->regs[VCPU_REGS_RCX], -1);
 
-               if (string_insn_completed(ctxt))
-                       ctxt->restart = false;
-               /*
-                * Re-enter guest when pio read ahead buffer is empty or,
-                * if it is not used, after each 1024 iteration.
-                */
-               else if ((r->end == 0 && !(c->regs[VCPU_REGS_RCX] & 0x3ff)) ||
-                        (r->end != 0 && r->end == r->pos)) {
-                       ctxt->restart = false;
-                       c->eip = ctxt->eip;
+               if (!string_insn_completed(ctxt)) {
+                       /*
+                        * Re-enter guest when pio read ahead buffer is empty
+                        * or, if it is not used, after each 1024 iteration.
+                        */
+                       if ((r->end != 0 || c->regs[VCPU_REGS_RCX] & 0x3ff) &&
+                           (r->end == 0 || r->end != r->pos)) {
+                               /*
+                                * Reset read cache. Usually happens before
+                                * decode, but since instruction is restarted
+                                * we have to do it here.
+                                */
+                               ctxt->decode.mem_read.end = 0;
+                               return EMULATION_RESTART;
+                       }
+                       goto done; /* skip rip writeback */
                }
        }
-       /*
-        * reset read cache here in case string instruction is restared
-        * without decoding
-        */
-       ctxt->decode.mem_read.end = 0;
-       if (!ctxt->restart)
-               ctxt->eip = c->eip;
+
+       ctxt->eip = c->eip;
 
 done:
-       return (rc == X86EMUL_UNHANDLEABLE) ? -1 : 0;
+       return (rc == X86EMUL_UNHANDLEABLE) ? EMULATION_FAILED : EMULATION_OK;
 
 twobyte_insn:
        switch (c->b) {
 
 restart:
        r = x86_emulate_insn(&vcpu->arch.emulate_ctxt);
 
-       if (r) { /* emulation failed */
+       if (r == EMULATION_FAILED) {
                if (reexecute_instruction(vcpu, cr2))
                        return EMULATE_DONE;
 
                return handle_emulation_failure(vcpu);
        }
 
-       r = EMULATE_DONE;
-
-       if (vcpu->arch.emulate_ctxt.exception >= 0)
+       if (vcpu->arch.emulate_ctxt.exception >= 0) {
                inject_emulated_exception(vcpu);
-       else if (vcpu->arch.pio.count) {
+               r = EMULATE_DONE;
+       } else if (vcpu->arch.pio.count) {
                if (!vcpu->arch.pio.in)
                        vcpu->arch.pio.count = 0;
                r = EMULATE_DO_MMIO;
                if (vcpu->mmio_is_write)
                        vcpu->mmio_needed = 0;
                r = EMULATE_DO_MMIO;
-       } else if (vcpu->arch.emulate_ctxt.restart)
+       } else if (r == EMULATION_RESTART)
                goto restart;
+       else
+               r = EMULATE_DONE;
 
        toggle_interruptibility(vcpu, vcpu->arch.emulate_ctxt.interruptibility);
        kvm_x86_ops->set_rflags(vcpu, vcpu->arch.emulate_ctxt.eflags);
        if (!irqchip_in_kernel(vcpu->kvm))
                kvm_set_cr8(vcpu, kvm_run->cr8);
 
-       if (vcpu->arch.pio.count || vcpu->mmio_needed ||
-           vcpu->arch.emulate_ctxt.restart) {
+       if (vcpu->arch.pio.count || vcpu->mmio_needed) {
                if (vcpu->mmio_needed) {
                        memcpy(vcpu->mmio_data, kvm_run->mmio.data, 8);
                        vcpu->mmio_read_completed = 1;