vhpt.c

xen虚拟机源代码安装包

/*
 * Initialize VHPT support.
 *
 * Copyright (C) 2004 Hewlett-Packard Co
 *	Dan Magenheimer <[e-mail]>
 *
 * Copyright (c) 2006 Isaku Yamahata <yamahata at valinux co jp>
 *                    VA Linux Systems Japan K.K.
 *                    per vcpu vhpt support
 */
#include <linux/config.h>
#include <linux/kernel.h>
#include <linux/init.h>

#include <asm/processor.h>
#include <asm/system.h>
#include <asm/pgalloc.h>
#include <asm/page.h>
#include <asm/vhpt.h>
#include <asm/vcpu.h>
#include <asm/vcpumask.h>
#include <asm/vmmu.h>

DEFINE_PER_CPU (unsigned long, vhpt_paddr);
DEFINE_PER_CPU (unsigned long, vhpt_pend);
#ifdef CONFIG_XEN_IA64_TLBFLUSH_CLOCK
DEFINE_PER_CPU(volatile u32, vhpt_tlbflush_timestamp);
#endif

static void
__vhpt_flush(unsigned long vhpt_maddr, unsigned long vhpt_size_log2)
{
	struct vhpt_lf_entry *v = (struct vhpt_lf_entry*)__va(vhpt_maddr);
	unsigned long num_entries = 1 << (vhpt_size_log2 - 5);
	int i;

	for (i = 0; i < num_entries; i++, v++)
		v->ti_tag = INVALID_TI_TAG;
}

void
local_vhpt_flush(void)
{
	/* increment flush clock before flush */
	u32 flush_time = tlbflush_clock_inc_and_return();
	__vhpt_flush(__ia64_per_cpu_var(vhpt_paddr), VHPT_SIZE_LOG2);
	/* this must be after flush */
	tlbflush_update_time(&__get_cpu_var(vhpt_tlbflush_timestamp),
	                     flush_time);
	perfc_incr(local_vhpt_flush);
}

void
vcpu_vhpt_flush(struct vcpu* v)
{
	unsigned long vhpt_size_log2 = VHPT_SIZE_LOG2;
#ifdef CONFIG_XEN_IA64_PERVCPU_VHPT
	if (HAS_PERVCPU_VHPT(v->domain))
		vhpt_size_log2 = v->arch.pta.size;
#endif
	__vhpt_flush(vcpu_vhpt_maddr(v), vhpt_size_log2);
	perfc_incr(vcpu_vhpt_flush);
}

static void
vhpt_erase(unsigned long vhpt_maddr, unsigned long vhpt_size_log2)
{
	struct vhpt_lf_entry *v = (struct vhpt_lf_entry*)__va(vhpt_maddr);
	unsigned long num_entries = 1 << (vhpt_size_log2 - 5);
	int i;

	for (i = 0; i < num_entries; i++, v++) {
		v->itir = 0;
		v->CChain = 0;
		v->page_flags = 0;
		v->ti_tag = INVALID_TI_TAG;
	}
	// initialize cache too???
}

void vhpt_insert (unsigned long vadr, unsigned long pte, unsigned long itir)
{
	struct vhpt_lf_entry *vlfe = (struct vhpt_lf_entry *)ia64_thash(vadr);
	unsigned long tag = ia64_ttag (vadr);

	/* Even though VHPT is per VCPU, still need to first disable the entry,
	 * because the processor may support speculative VHPT walk.  */
	vlfe->ti_tag = INVALID_TI_TAG;
	wmb();
	vlfe->itir = itir;
	vlfe->page_flags = pte | _PAGE_P;
	*(volatile unsigned long*)&vlfe->ti_tag = tag;
}

void vhpt_multiple_insert(unsigned long vaddr, unsigned long pte,
			   unsigned long itir)
{
	unsigned char ps = current->arch.vhpt_pg_shift;
	ia64_itir_t _itir = {.itir = itir};
	unsigned long mask = (1L << _itir.ps) - 1;
	int i;

	if (_itir.ps - ps > 10 && !running_on_sim) {
		// if this happens, we may want to revisit this algorithm
		panic("vhpt_multiple_insert:logps-PAGE_SHIFT>10,spinning..\n");
	}
	if (_itir.ps - ps > 2) {
		// FIXME: Should add counter here to see how often this
		//  happens (e.g. for 16MB pages!) and determine if it
		//  is a performance problem.  On a quick look, it takes
		//  about 39000 instrs for a 16MB page and it seems to occur
		//  only a few times/second, so OK for now.
		//  An alternate solution would be to just insert the one
		//  16KB in the vhpt (but with the full mapping)?
		//printk("vhpt_multiple_insert: logps-PAGE_SHIFT==%d,"
			//"va=%p, pa=%p, pa-masked=%p\n",
			//logps-PAGE_SHIFT,vaddr,pte&_PFN_MASK,
			//(pte&_PFN_MASK)&~mask);
	}
	vaddr &= ~mask;
	pte = ((pte & _PFN_MASK) & ~mask) | (pte & ~_PFN_MASK);
	for (i = 1L << (_itir.ps - ps); i > 0; i--) {
		vhpt_insert(vaddr, pte, _itir.itir);
		vaddr += (1L << ps);
	}
}

void __init vhpt_init(void)
{
	unsigned long paddr;
	struct page_info *page;
#if !VHPT_ENABLED
	return;
#endif
	/* This allocation only holds true if vhpt table is unique for
	 * all domains. Or else later new vhpt table should be allocated
	 * from domain heap when each domain is created. Assume xen buddy
	 * allocator can provide natural aligned page by order?
	 */
	page = alloc_domheap_pages(NULL, VHPT_SIZE_LOG2 - PAGE_SHIFT, 0);
	if (!page)
		panic("vhpt_init: can't allocate VHPT!\n");
	paddr = page_to_maddr(page);
	if (paddr & ((1 << VHPT_SIZE_LOG2) - 1))
		panic("vhpt_init: bad VHPT alignment!\n");
	__get_cpu_var(vhpt_paddr) = paddr;
	__get_cpu_var(vhpt_pend) = paddr + (1 << VHPT_SIZE_LOG2) - 1;
	printk(XENLOG_DEBUG "vhpt_init: vhpt paddr=0x%lx, end=0x%lx\n",
	       paddr, __get_cpu_var(vhpt_pend));
	vhpt_erase(paddr, VHPT_SIZE_LOG2);
	// we don't enable VHPT here.
	// context_switch() or schedule_tail() does it.
}

#ifdef CONFIG_XEN_IA64_PERVCPU_VHPT
void
domain_set_vhpt_size(struct domain *d, int8_t vhpt_size_log2)
{
	if (vhpt_size_log2 == -1) {
		d->arch.has_pervcpu_vhpt = 0;
		printk(XENLOG_INFO "XEN_DOMCTL_arch_setup: "
		       "domain %d VHPT is global.\n", d->domain_id);
	} else {
		d->arch.has_pervcpu_vhpt = 1;
		d->arch.vhpt_size_log2 = vhpt_size_log2;
		printk(XENLOG_INFO "XEN_DOMCTL_arch_setup: "
		       "domain %d VHPT is per vcpu. size=2**%d\n",
		       d->domain_id, vhpt_size_log2);
	}
}

int
pervcpu_vhpt_alloc(struct vcpu *v)
{
	unsigned long vhpt_size_log2 = VHPT_SIZE_LOG2;

	if (v->domain->arch.vhpt_size_log2 > 0)
	    vhpt_size_log2 =
		canonicalize_vhpt_size(v->domain->arch.vhpt_size_log2);
	printk(XENLOG_DEBUG "%s vhpt_size_log2=%ld\n",
	       __func__, vhpt_size_log2);
	v->arch.vhpt_entries =
		(1UL << vhpt_size_log2) / sizeof(struct vhpt_lf_entry);
	v->arch.vhpt_page =
		alloc_domheap_pages(NULL, vhpt_size_log2 - PAGE_SHIFT, 0);
	if (!v->arch.vhpt_page)
		return -ENOMEM;
	
	v->arch.vhpt_maddr = page_to_maddr(v->arch.vhpt_page);
	if (v->arch.vhpt_maddr & ((1 << VHPT_SIZE_LOG2) - 1))
		panic("pervcpu_vhpt_init: bad VHPT alignment!\n");

	v->arch.pta.val = 0; // to zero reserved bits
	v->arch.pta.ve = 1; // enable vhpt
	v->arch.pta.size = vhpt_size_log2;
	v->arch.pta.vf = 1; // long format
	v->arch.pta.base = __va_ul(v->arch.vhpt_maddr) >> 15;

	vhpt_erase(v->arch.vhpt_maddr, vhpt_size_log2);
	smp_mb(); // per vcpu vhpt may be used by another physical cpu.
	return 0;
}

void
pervcpu_vhpt_free(struct vcpu *v)
{
	if (likely(v->arch.vhpt_page != NULL))
		free_domheap_pages(v->arch.vhpt_page,
		                   v->arch.pta.size - PAGE_SHIFT);
}
#endif

void
domain_purge_swtc_entries(struct domain *d)
{
	struct vcpu* v;
	for_each_vcpu(d, v) {
		if (!v->is_initialised)
			continue;

		/* Purge TC entries.
		   FIXME: clear only if match.  */
		vcpu_purge_tr_entry(&PSCBX(v,dtlb));
		vcpu_purge_tr_entry(&PSCBX(v,itlb));
	}
}

void
domain_purge_swtc_entries_vcpu_dirty_mask(struct domain* d,
                                          vcpumask_t vcpu_dirty_mask)
{
	int vcpu;

	for_each_vcpu_mask(vcpu, vcpu_dirty_mask) {
		struct vcpu* v = d->vcpu[vcpu];
		if (!v->is_initialised)
			continue;

		/* Purge TC entries.
		   FIXME: clear only if match.  */
		vcpu_purge_tr_entry(&PSCBX(v, dtlb));
		vcpu_purge_tr_entry(&PSCBX(v, itlb));
	}
}

// SMP: we can't assume v == current, vcpu might move to another physical cpu.
// So memory barrier is necessary.
// if we can guranttee that vcpu can run on only this physical cpu
// (e.g. vcpu == current), smp_mb() is unnecessary.
void vcpu_flush_vtlb_all(struct vcpu *v)
{
	/* First VCPU tlb.  */
	vcpu_purge_tr_entry(&PSCBX(v,dtlb));
	vcpu_purge_tr_entry(&PSCBX(v,itlb));
	smp_mb();

	/* Then VHPT.  */
	if (HAS_PERVCPU_VHPT(v->domain))
		vcpu_vhpt_flush(v);
	else
		local_vhpt_flush();
	smp_mb();

	/* Then mTLB.  */
	local_flush_tlb_all();

	/* We could clear bit in d->domain_dirty_cpumask only if domain d in
	   not running on this processor.  There is currently no easy way to
	   check this.  */

	perfc_incr(vcpu_flush_vtlb_all);
}

static void __vcpu_flush_vtlb_all(void *vcpu)
{
	vcpu_flush_vtlb_all((struct vcpu*)vcpu);
}

// caller must incremented reference count to d somehow.
void domain_flush_vtlb_all(struct domain* d)
{
	int cpu = smp_processor_id ();
	struct vcpu *v;

	for_each_vcpu(d, v) {
		if (!v->is_initialised)
			continue;

		if (VMX_DOMAIN(v)) {
			// This code may be called for remapping shared_info
			// and grant_table from guest_physmap_remove_page()
			// in arch_memory_op() XENMEM_add_to_physmap to realize
			// PV-on-HVM feature.
			vmx_vcpu_flush_vtlb_all(v);