pmap.c

早期freebsd实现

	vm_offset_t	start;
	vm_offset_t	end;
	int		prot;
{
#ifdef DEBUG
	if (pmapdebug & PDB_FOLLOW)
		printf("pmap_map(%x, %x, %x, %x)\n", virt, start, end, prot);
#endif
	while (start < end) {
		pmap_enter(kernel_pmap, virt, start, prot, FALSE);
		virt += PAGE_SIZE;
		start += PAGE_SIZE;
	}
	return(virt);
}

/*
 *	Create and return a physical map.
 *
 *	If the size specified for the map
 *	is zero, the map is an actual physical
 *	map, and may be referenced by the
 *	hardware.
 *
 *	If the size specified is non-zero,
 *	the map will be used in software only, and
 *	is bounded by that size.
 */
pmap_t
pmap_create(size)
	vm_size_t	size;
{
	register pmap_t pmap;

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_CREATE))
		printf("pmap_create(%x)\n", size);
#endif
	/*
	 * Software use map does not need a pmap
	 */
	if (size)
		return(NULL);

	/* XXX: is it ok to wait here? */
	pmap = (pmap_t) malloc(sizeof *pmap, M_VMPMAP, M_WAITOK);
#ifdef notifwewait
	if (pmap == NULL)
		panic("pmap_create: cannot allocate a pmap");
#endif
	bzero(pmap, sizeof(*pmap));
	pmap_pinit(pmap);
	return (pmap);
}

/*
 * Initialize a preallocated and zeroed pmap structure,
 * such as one in a vmspace structure.
 */
void
pmap_pinit(pmap)
	register struct pmap *pmap;
{

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_CREATE))
		printf("pmap_pinit(%x)\n", pmap);
#endif
	/*
	 * No need to allocate page table space yet but we do need a
	 * valid segment table.  Initially, we point everyone at the
	 * "null" segment table.  On the first pmap_enter, a real
	 * segment table will be allocated.
	 */
	pmap->pm_stab = Segtabzero;
	pmap->pm_stpa = Segtabzeropa;
#if defined(LUNA2)
	if (mmutype == MMU_68040)
		pmap->pm_stfree = protostfree;
#endif
	pmap->pm_stchanged = TRUE;
	pmap->pm_count = 1;
	simple_lock_init(&pmap->pm_lock);
}

/*
 *	Retire the given physical map from service.
 *	Should only be called if the map contains
 *	no valid mappings.
 */
void
pmap_destroy(pmap)
	register pmap_t pmap;
{
	int count;

#ifdef DEBUG
	if (pmapdebug & PDB_FOLLOW)
		printf("pmap_destroy(%x)\n", pmap);
#endif
	if (pmap == NULL)
		return;

	simple_lock(&pmap->pm_lock);
	count = --pmap->pm_count;
	simple_unlock(&pmap->pm_lock);
	if (count == 0) {
		pmap_release(pmap);
		free((caddr_t)pmap, M_VMPMAP);
	}
}

/*
 * Release any resources held by the given physical map.
 * Called when a pmap initialized by pmap_pinit is being released.
 * Should only be called if the map contains no valid mappings.
 */
void
pmap_release(pmap)
	register struct pmap *pmap;
{

#ifdef DEBUG
	if (pmapdebug & PDB_FOLLOW)
		printf("pmap_release(%x)\n", pmap);
#endif
#ifdef notdef /* DIAGNOSTIC */
	/* count would be 0 from pmap_destroy... */
	simple_lock(&pmap->pm_lock);
	if (pmap->pm_count != 1)
		panic("pmap_release count");
#endif
	if (pmap->pm_ptab)
		kmem_free_wakeup(pt_map, (vm_offset_t)pmap->pm_ptab,
				 LUNA_MAX_PTSIZE);
	if (pmap->pm_stab != Segtabzero)
		kmem_free(kernel_map, (vm_offset_t)pmap->pm_stab, LUNA_STSIZE);
}

/*
 *	Add a reference to the specified pmap.
 */
void
pmap_reference(pmap)
	pmap_t	pmap;
{
#ifdef DEBUG
	if (pmapdebug & PDB_FOLLOW)
		printf("pmap_reference(%x)\n", pmap);
#endif
	if (pmap != NULL) {
		simple_lock(&pmap->pm_lock);
		pmap->pm_count++;
		simple_unlock(&pmap->pm_lock);
	}
}

/*
 *	Remove the given range of addresses from the specified map.
 *
 *	It is assumed that the start and end are properly
 *	rounded to the page size.
 */
void
pmap_remove(pmap, sva, eva)
	register pmap_t pmap;
	register vm_offset_t sva, eva;
{
	register vm_offset_t nssva;
	register pt_entry_t *pte;
	boolean_t firstpage, needcflush;
	int flags;

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_REMOVE|PDB_PROTECT))
		printf("pmap_remove(%x, %x, %x)\n", pmap, sva, eva);
#endif

	if (pmap == NULL)
		return;

#ifdef PMAPSTATS
	remove_stats.calls++;
#endif
	firstpage = TRUE;
	needcflush = FALSE;
	flags = active_pmap(pmap) ? PRM_TFLUSH : 0;
	while (sva < eva) {
		nssva = luna_trunc_seg(sva) + LUNA_SEG_SIZE;
		if (nssva == 0 || nssva > eva)
			nssva = eva;
		/*
		 * If VA belongs to an unallocated segment,
		 * skip to the next segment boundary.
		 */
		if (!pmap_ste_v(pmap, sva)) {
			sva = nssva;
			continue;
		}
		/*
		 * Invalidate every valid mapping within this segment.
		 */
		pte = pmap_pte(pmap, sva);
		while (sva < nssva) {
			if (pmap_pte_v(pte)) {
#ifdef HAVEVAC
				if (pmap_aliasmask) {
					/*
					 * Purge kernel side of VAC to ensure
					 * we get the correct state of any
					 * hardware maintained bits.
					 */
					if (firstpage) {
						DCIS();
#ifdef PMAPSTATS
						remove_stats.sflushes++;
#endif
					}
					/*
					 * Remember if we may need to
					 * flush the VAC due to a non-CI
					 * mapping.
					 */
					if (!needcflush && !pmap_pte_ci(pte))
						needcflush = TRUE;

				}
#endif
				pmap_remove_mapping(pmap, sva, pte, flags);
				firstpage = FALSE;
			}
			pte++;
			sva += PAGE_SIZE;
		}
	}
	/*
	 * Didn't do anything, no need for cache flushes
	 */
	if (firstpage)
		return;
#ifdef HAVEVAC
	/*
	 * In a couple of cases, we don't need to worry about flushing
	 * the VAC:
	 * 	1. if this is a kernel mapping,
	 *	   we have already done it
	 *	2. if it is a user mapping not for the current process,
	 *	   it won't be there
	 */
	if (pmap_aliasmask &&
	    (pmap == kernel_pmap || pmap != curproc->p_vmspace->vm_map.pmap))
		needcflush = FALSE;
#ifdef DEBUG
	if (pmap_aliasmask && (pmapvacflush & PVF_REMOVE)) {
		if (pmapvacflush & PVF_TOTAL)
			DCIA();
		else if (pmap == kernel_pmap)
			DCIS();
		else
			DCIU();
	} else
#endif
	if (needcflush) {
		if (pmap == kernel_pmap) {
			DCIS();
#ifdef PMAPSTATS
			remove_stats.sflushes++;
#endif
		} else {
			DCIU();
#ifdef PMAPSTATS
			remove_stats.uflushes++;
#endif
		}
	}
#endif
}

/*
 *	pmap_page_protect:
 *
 *	Lower the permission for all mappings to a given page.
 */
void
pmap_page_protect(pa, prot)
	vm_offset_t	pa;
	vm_prot_t	prot;
{
	register pv_entry_t pv;
	int s;

#ifdef DEBUG
	if ((pmapdebug & (PDB_FOLLOW|PDB_PROTECT)) ||
	    prot == VM_PROT_NONE && (pmapdebug & PDB_REMOVE))
		printf("pmap_page_protect(%x, %x)\n", pa, prot);
#endif
	if (pa < vm_first_phys || pa >= vm_last_phys)
		return;

	switch (prot) {
	case VM_PROT_READ|VM_PROT_WRITE:
	case VM_PROT_ALL:
		break;
	/* copy_on_write */
	case VM_PROT_READ:
	case VM_PROT_READ|VM_PROT_EXECUTE:
		pmap_changebit(pa, PG_RO, TRUE);
		break;
	/* remove_all */
	default:
		pv = pa_to_pvh(pa);
		s = splimp();
		while (pv->pv_pmap != NULL) {
#ifdef DEBUG
			if (!pmap_ste_v(pv->pv_pmap, pv->pv_va) ||
			    pmap_pte_pa(pmap_pte(pv->pv_pmap,pv->pv_va)) != pa)
				panic("pmap_page_protect: bad mapping");
#endif
			pmap_remove_mapping(pv->pv_pmap, pv->pv_va,
					    PT_ENTRY_NULL,
					    PRM_TFLUSH|PRM_CFLUSH);
		}
		splx(s);
		break;
	}
}

/*
 *	Set the physical protection on the
 *	specified range of this map as requested.
 */
void
pmap_protect(pmap, sva, eva, prot)
	register pmap_t	pmap;
	register vm_offset_t sva, eva;
	vm_prot_t prot;
{
	register vm_offset_t nssva;
	register pt_entry_t *pte;
	boolean_t firstpage, needtflush;
	int isro;

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_PROTECT))
		printf("pmap_protect(%x, %x, %x, %x)\n", pmap, sva, eva, prot);
#endif

	if (pmap == NULL)
		return;

#ifdef PMAPSTATS
	protect_stats.calls++;
#endif
	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
		pmap_remove(pmap, sva, eva);
		return;
	}
	if (prot & VM_PROT_WRITE)
		return;

	isro = pte_prot(pmap, prot);
	needtflush = active_pmap(pmap);
	firstpage = TRUE;
	while (sva < eva) {
		nssva = luna_trunc_seg(sva) + LUNA_SEG_SIZE;
		if (nssva == 0 || nssva > eva)
			nssva = eva;
		/*
		 * If VA belongs to an unallocated segment,
		 * skip to the next segment boundary.
		 */
		if (!pmap_ste_v(pmap, sva)) {
			sva = nssva;
			continue;
		}
		/*
		 * Change protection on mapping if it is valid and doesn't
		 * already have the correct protection.
		 */
		pte = pmap_pte(pmap, sva);
		while (sva < nssva) {
			if (pmap_pte_v(pte) && pmap_pte_prot_chg(pte, isro)) {
#ifdef HAVEVAC
				/*
				 * Purge kernel side of VAC to ensure we
				 * get the correct state of any hardware
				 * maintained bits.
				 *
				 * XXX do we need to clear the VAC in
				 * general to reflect the new protection?
				 */
				if (firstpage && pmap_aliasmask)
					DCIS();
#endif
#if defined(LUNA2)
				/*
				 * Clear caches if making RO (see section
				 * "7.3 Cache Coherency" in the manual).
				 */
				if (isro && mmutype == MMU_68040) {
					vm_offset_t pa = pmap_pte_pa(pte);

					DCFP(pa);
					ICPP(pa);
				}
#endif
				pmap_pte_set_prot(pte, isro);
				if (needtflush)
					TBIS(sva);
#ifdef PMAPSTATS
				protect_stats.changed++;
#endif
				firstpage = FALSE;
			}
#ifdef PMAPSTATS
			else if (pmap_pte_v(pte)) {
				if (isro)
					protect_stats.alreadyro++;
				else
					protect_stats.alreadyrw++;
			}
#endif
			pte++;
			sva += PAGE_SIZE;
		}
	}
#if defined(HAVEVAC) && defined(DEBUG)
	if (pmap_aliasmask && (pmapvacflush & PVF_PROTECT)) {
		if (pmapvacflush & PVF_TOTAL)
			DCIA();
		else if (pmap == kernel_pmap)
			DCIS();
		else
			DCIU();
	}
#endif
}

/*
 *	Insert the given physical page (p) at
 *	the specified virtual address (v) in the
 *	target physical map with the protection requested.
 *
 *	If specified, the page will be wired down, meaning
 *	that the related pte can not be reclaimed.
 *
 *	NB:  This is the only routine which MAY NOT lazy-evaluate
 *	or lose information.  That is, this routine must actually
 *	insert this page into the given map NOW.
 */
void
pmap_enter(pmap, va, pa, prot, wired)
	register pmap_t pmap;
	vm_offset_t va;
	register vm_offset_t pa;
	vm_prot_t prot;
	boolean_t wired;
{
	register pt_entry_t *pte;
	register int npte;
	vm_offset_t opa;
	boolean_t cacheable = TRUE;
	boolean_t checkpv = TRUE;

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_ENTER))
		printf("pmap_enter(%x, %x, %x, %x, %x)\n",
		       pmap, va, pa, prot, wired);
#endif
	if (pmap == NULL)
		return;

#ifdef PMAPSTATS
	if (pmap == kernel_pmap)
		enter_stats.kernel++;
	else
		enter_stats.user++;
#endif
	/*
	 * For user mapping, allocate kernel VM resources if necessary.
	 */
	if (pmap->pm_ptab == NULL)
		pmap->pm_ptab = (pt_entry_t *)
			kmem_alloc_wait(pt_map, LUNA_MAX_PTSIZE);

	/*
	 * Segment table entry not valid, we need a new PT page
	 */
	if (!pmap_ste_v(pmap, va))
		pmap_enter_ptpage(pmap, va);

	pa = luna_trunc_page(pa);