vmx_init.c

xen虚拟机源代码安装包

	BUG_ON(v != current);
	
	ia64_call_vsa(PAL_VPS_SAVE, (u64)v->arch.privregs, 1, 0, 0, 0, 0, 0);

	/* Need to save KR when domain switch, though HV itself doesn;t
	 * use them.
	 */
	v->arch.arch_vmx.vkr[0] = ia64_get_kr(0);
	v->arch.arch_vmx.vkr[1] = ia64_get_kr(1);
	v->arch.arch_vmx.vkr[2] = ia64_get_kr(2);
	v->arch.arch_vmx.vkr[3] = ia64_get_kr(3);
	v->arch.arch_vmx.vkr[4] = ia64_get_kr(4);
	v->arch.arch_vmx.vkr[5] = ia64_get_kr(5);
	v->arch.arch_vmx.vkr[6] = ia64_get_kr(6);
	v->arch.arch_vmx.vkr[7] = ia64_get_kr(7);
}

/* Even guest is in physical mode, we still need such double mapping */
void
vmx_load_state(struct vcpu *v)
{
	BUG_ON(v != current);

	vmx_load_all_rr(v);

	/* vmx_load_all_rr() pins down v->arch.privregs with both dtr/itr*/
	ia64_call_vsa(PAL_VPS_RESTORE, (u64)v->arch.privregs, 1, 0, 0, 0, 0, 0);

	ia64_set_kr(0, v->arch.arch_vmx.vkr[0]);
	ia64_set_kr(1, v->arch.arch_vmx.vkr[1]);
	ia64_set_kr(2, v->arch.arch_vmx.vkr[2]);
	ia64_set_kr(3, v->arch.arch_vmx.vkr[3]);
	ia64_set_kr(4, v->arch.arch_vmx.vkr[4]);
	ia64_set_kr(5, v->arch.arch_vmx.vkr[5]);
	ia64_set_kr(6, v->arch.arch_vmx.vkr[6]);
	ia64_set_kr(7, v->arch.arch_vmx.vkr[7]);
	/* Guest vTLB is not required to be switched explicitly, since
	 * anchored in vcpu */

	migrate_timer(&v->arch.arch_vmx.vtm.vtm_timer, v->processor);
}

static int
vmx_vcpu_initialise(struct vcpu *v)
{
	struct vmx_ioreq_page *iorp = &v->domain->arch.hvm_domain.ioreq;

	int rc = alloc_unbound_xen_event_channel(v, 0);
	if (rc < 0)
		return rc;
	v->arch.arch_vmx.xen_port = rc;

	spin_lock(&iorp->lock);
	if (v->domain->arch.vmx_platform.ioreq.va != 0) {
		vcpu_iodata_t *p = get_vio(v);
		p->vp_eport = v->arch.arch_vmx.xen_port;
	}
	spin_unlock(&iorp->lock);

	gdprintk(XENLOG_INFO, "Allocated port %ld for hvm %d vcpu %d.\n",
		 v->arch.arch_vmx.xen_port, v->domain->domain_id, v->vcpu_id);

	return 0;
}

static int vmx_create_event_channels(struct vcpu *v)
{
	struct vcpu *o;

	if (v->vcpu_id == 0) {
		/* Ugly: create event channels for every vcpu when vcpu 0
		   starts, so that they're available for ioemu to bind to. */
		for_each_vcpu(v->domain, o) {
			int rc = vmx_vcpu_initialise(o);
			if (rc < 0) //XXX error recovery
				return rc;
		}
	}

	return 0;
}

/*
 * Event channel has destoryed in domain_kill(), so we needn't
 * do anything here
 */
static void vmx_release_assist_channel(struct vcpu *v)
{
	return;
}

/* following three functions are based from hvm_xxx_ioreq_page()
 * in xen/arch/x86/hvm/hvm.c */
static void vmx_init_ioreq_page(
	struct domain *d, struct vmx_ioreq_page *iorp)
{
	memset(iorp, 0, sizeof(*iorp));
	spin_lock_init(&iorp->lock);
	domain_pause(d);
}

static void vmx_destroy_ioreq_page(
	struct domain *d, struct vmx_ioreq_page *iorp)
{
	spin_lock(&iorp->lock);

	ASSERT(d->is_dying);

	if (iorp->va != NULL) {
		put_page(iorp->page);
		iorp->page = NULL;
		iorp->va = NULL;
	}

	spin_unlock(&iorp->lock);
}

int vmx_set_ioreq_page(
	struct domain *d, struct vmx_ioreq_page *iorp, unsigned long gpfn)
{
	struct page_info *page;
	unsigned long mfn;
	pte_t pte;

	pte = *lookup_noalloc_domain_pte(d, gpfn << PAGE_SHIFT);
	if (!pte_present(pte) || !pte_mem(pte))
		return -EINVAL;
	mfn = (pte_val(pte) & _PFN_MASK) >> PAGE_SHIFT;
	ASSERT(mfn_valid(mfn));

	page = mfn_to_page(mfn);
	if (get_page(page, d) == 0)
		return -EINVAL;

	spin_lock(&iorp->lock);

	if ((iorp->va != NULL) || d->is_dying) {
		spin_unlock(&iorp->lock);
		put_page(page);
		return -EINVAL;
	}

	iorp->va = mfn_to_virt(mfn);
	iorp->page = page;

	spin_unlock(&iorp->lock);

	domain_unpause(d);

	return 0;
}

/*
 * Initialize VMX envirenment for guest. Only the 1st vp/vcpu
 * is registered here.
 */
int
vmx_final_setup_guest(struct vcpu *v)
{
	vpd_t *vpd;
	int rc;

	vpd = alloc_vpd();
	ASSERT(vpd);
	if (!vpd)
		return -ENOMEM;

	v->arch.privregs = (mapped_regs_t *)vpd;
	vpd->vpd_low.virt_env_vaddr = vm_buffer;
    
	/* Per-domain vTLB and vhpt implementation. Now vmx domain will stick
	 * to this solution. Maybe it can be deferred until we know created
	 * one as vmx domain */
	rc = init_domain_tlb(v);
	if (rc)
		return rc;

	if (!v->domain->arch.is_sioemu) {
		rc = vmx_create_event_channels(v);
		if (rc)
			return rc;
	}

	/* v->arch.schedule_tail = arch_vmx_do_launch; */
	vmx_create_vp(v);

	/* Physical mode emulation initialization, including
	* emulation ID allcation and related memory request
	*/
	physical_mode_init(v);

	vlsapic_reset(v);
	vtm_init(v);

	/* Set up guest 's indicator for VTi domain*/
	set_bit(ARCH_VMX_DOMAIN, &v->arch.arch_vmx.flags);

	return 0;
}

void
vmx_relinquish_guest_resources(struct domain *d)
{
	struct vcpu *v;

	if (d->arch.is_sioemu)
		return;

	for_each_vcpu(d, v)
		vmx_release_assist_channel(v);

	vacpi_relinquish_resources(d);

	vmx_destroy_ioreq_page(d, &d->arch.vmx_platform.ioreq);
	vmx_destroy_ioreq_page(d, &d->arch.vmx_platform.buf_ioreq);
	vmx_destroy_ioreq_page(d, &d->arch.vmx_platform.buf_pioreq);
}

void
vmx_relinquish_vcpu_resources(struct vcpu *v)
{
	vtime_t *vtm = &(v->arch.arch_vmx.vtm);

	kill_timer(&vtm->vtm_timer);

	if (v->arch.arch_vmx.sioemu_info_mva)
		put_page(virt_to_page((unsigned long)
		                      v->arch.arch_vmx.sioemu_info_mva));

	free_domain_tlb(v);
	free_vpd(v);
}

typedef struct io_range {
	unsigned long start;
	unsigned long size;
	unsigned long type;
} io_range_t;

static const io_range_t io_ranges[] = {
	{VGA_IO_START, VGA_IO_SIZE, GPFN_FRAME_BUFFER << PAGE_SHIFT},
	{MMIO_START, MMIO_SIZE, GPFN_LOW_MMIO << PAGE_SHIFT},
	{LEGACY_IO_START, LEGACY_IO_SIZE, GPFN_LEGACY_IO << PAGE_SHIFT},
	{IO_SAPIC_START, IO_SAPIC_SIZE, GPFN_IOSAPIC << PAGE_SHIFT},
	{PIB_START, PIB_SIZE, GPFN_PIB << PAGE_SHIFT},
};

// The P2M table is built in libxc/ia64/xc_ia64_hvm_build.c @ setup_guest()
// so only mark IO memory space here
static void vmx_build_io_physmap_table(struct domain *d)
{
	unsigned long i, j;

	/* Mark I/O ranges */
	for (i = 0; i < (sizeof(io_ranges) / sizeof(io_range_t)); i++) {
		for (j = io_ranges[i].start;
		     j < io_ranges[i].start + io_ranges[i].size; j += PAGE_SIZE)
			(void)__assign_domain_page(d, j, io_ranges[i].type,
			                           ASSIGN_writable | ASSIGN_io);
	}

}

int vmx_setup_platform(struct domain *d)
{
	ASSERT(d != dom0); /* only for non-privileged vti domain */

	if (!d->arch.is_sioemu) {
		vmx_build_io_physmap_table(d);

		vmx_init_ioreq_page(d, &d->arch.vmx_platform.ioreq);
		vmx_init_ioreq_page(d, &d->arch.vmx_platform.buf_ioreq);
		vmx_init_ioreq_page(d, &d->arch.vmx_platform.buf_pioreq);
	}
	/* TEMP */
	d->arch.vmx_platform.pib_base = 0xfee00000UL;

	d->arch.sal_data = xmalloc(struct xen_sal_data);
	if (d->arch.sal_data == NULL)
		return -ENOMEM;

	/* Only open one port for I/O and interrupt emulation */
	memset(&d->shared_info->evtchn_mask[0], 0xff,
	       sizeof(d->shared_info->evtchn_mask));

	/* Initialize iosapic model within hypervisor */
	viosapic_init(d);

	if (!d->arch.is_sioemu)
		vacpi_init(d);

	if (d->arch.is_sioemu) {
		int i;
		for (i = 1; i < MAX_VIRT_CPUS; i++)
			d->shared_info->vcpu_info[i].evtchn_upcall_mask = 1;
	}

	return 0;
}

void vmx_do_resume(struct vcpu *v)
{
	ioreq_t *p;

	vmx_load_state(v);

	if (v->domain->arch.is_sioemu)
		return;

	/* stolen from hvm_do_resume() in arch/x86/hvm/hvm.c */
	/* NB. Optimised for common case (p->state == STATE_IOREQ_NONE). */
	p = &get_vio(v)->vp_ioreq;
	while (p->state != STATE_IOREQ_NONE) {
		switch (p->state) {
		case STATE_IORESP_READY: /* IORESP_READY -> NONE */
			vmx_io_assist(v);
			break;
		case STATE_IOREQ_READY:
		case STATE_IOREQ_INPROCESS:
			/* IOREQ_{READY,INPROCESS} -> IORESP_READY */
			wait_on_xen_event_channel(v->arch.arch_vmx.xen_port,
					  (p->state != STATE_IOREQ_READY) &&
					  (p->state != STATE_IOREQ_INPROCESS));
			break;
		default:
			gdprintk(XENLOG_ERR,
				 "Weird HVM iorequest state %d.\n", p->state);
			domain_crash_synchronous();
		}
	}
}