pmap.c

早期freebsd实现

 *
 *	It is assumed that the start and end are properly
 *	rounded to the page size.
 */
void
pmap_remove(pmap, sva, eva)
	register struct pmap *pmap;
	vm_offset_t sva, eva;
{
	register vm_offset_t pa, va;
	register pt_entry_t *pte;
	register pv_entry_t pv, npv;
	register int ix;
	pmap_t ptpmap;
	int *pde, s, bits;
	boolean_t firstpage = TRUE;
	boolean_t flushcache = FALSE;
#ifdef DEBUG
	pt_entry_t opte;

	if (pmapdebug & (PDB_FOLLOW|PDB_REMOVE|PDB_PROTECT))
		printf("pmap_remove(%x, %x, %x)", pmap, sva, eva);
	if (eva >= USRSTACK && eva <= UPT_MAX_ADDRESS)
		nullop();
#endif

	if (pmap == NULL)
		return;

#ifdef DEBUG
	remove_stats.calls++;
#endif
	for (va = sva; va < eva; va += PAGE_SIZE) {
		/*
		 * Weed out invalid mappings.
		 * Note: we assume that the page directory table is
	 	 * always allocated, and in kernel virtual.
		 */
		if (!pmap_pde_v(pmap_pde(pmap, va)))
			continue;

		pte = pmap_pte(pmap, va);
		if (pte == 0)
			continue;
		pa = pmap_pte_pa(pte);
		if (pa == 0)
			continue;
#ifdef DEBUG
		opte = *pte;
		remove_stats.removes++;
#endif
		/*
		 * Update statistics
		 */
		if (pmap_pte_w(pte))
			pmap->pm_stats.wired_count--;
		pmap->pm_stats.resident_count--;

		/*
		 * Invalidate the PTEs.
		 * XXX: should cluster them up and invalidate as many
		 * as possible at once.
		 */
#ifdef DEBUG
		if (pmapdebug & PDB_REMOVE)
			printf("remove: inv %x ptes at %x(%x) ",
			       i386pagesperpage, pte, *(int *)pte);
#endif
		bits = ix = 0;
		do {
			bits |= *(int *)pte & (PG_U|PG_M);
			*(int *)pte++ = 0;
			/*TBIS(va + ix * I386_PAGE_SIZE);*/
		} while (++ix != i386pagesperpage);
		if (pmap == &curproc->p_vmspace->vm_pmap)
			pmap_activate(pmap, (struct pcb *)curproc->p_addr);
		/* are we current address space or kernel? */
		/*if (pmap->pm_pdir[PTDPTDI].pd_pfnum == PTDpde.pd_pfnum
			|| pmap == kernel_pmap)
		load_cr3(curpcb->pcb_ptd);*/
		tlbflush();

#ifdef needednotdone
reduce wiring count on page table pages as references drop
#endif

		/*
		 * Remove from the PV table (raise IPL since we
		 * may be called at interrupt time).
		 */
		if (pa < vm_first_phys || pa >= vm_last_phys)
			continue;
		pv = pa_to_pvh(pa);
		s = splimp();
		/*
		 * If it is the first entry on the list, it is actually
		 * in the header and we must copy the following entry up
		 * to the header.  Otherwise we must search the list for
		 * the entry.  In either case we free the now unused entry.
		 */
		if (pmap == pv->pv_pmap && va == pv->pv_va) {
			npv = pv->pv_next;
			if (npv) {
				*pv = *npv;
				free((caddr_t)npv, M_VMPVENT);
			} else
				pv->pv_pmap = NULL;
#ifdef DEBUG
			remove_stats.pvfirst++;
#endif
		} else {
			for (npv = pv->pv_next; npv; npv = npv->pv_next) {
#ifdef DEBUG
				remove_stats.pvsearch++;
#endif
				if (pmap == npv->pv_pmap && va == npv->pv_va)
					break;
				pv = npv;
			}
#ifdef DEBUG
			if (npv == NULL)
				panic("pmap_remove: PA not in pv_tab");
#endif
			pv->pv_next = npv->pv_next;
			free((caddr_t)npv, M_VMPVENT);
			pv = pa_to_pvh(pa);
		}

#ifdef notdef
[tally number of pagetable pages, if sharing of ptpages adjust here]
#endif
		/*
		 * Update saved attributes for managed page
		 */
		pmap_attributes[pa_index(pa)] |= bits;
		splx(s);
	}
#ifdef notdef
[cache and tlb flushing, if needed]
#endif
}

/*
 *	Routine:	pmap_remove_all
 *	Function:
 *		Removes this physical page from
 *		all physical maps in which it resides.
 *		Reflects back modify bits to the pager.
 */
void
pmap_remove_all(pa)
	vm_offset_t pa;
{
	register pv_entry_t pv;
	int s;

#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_REMOVE|PDB_PROTECT))
		printf("pmap_remove_all(%x)", pa);
	/*pmap_pvdump(pa);*/
#endif
	/*
	 * Not one of ours
	 */
	if (pa < vm_first_phys || pa >= vm_last_phys)
		return;

	pv = pa_to_pvh(pa);
	s = splimp();
	/*
	 * Do it the easy way for now
	 */
	while (pv->pv_pmap != NULL) {
#ifdef DEBUG
		if (!pmap_pde_v(pmap_pde(pv->pv_pmap, pv->pv_va)) ||
		    pmap_pte_pa(pmap_pte(pv->pv_pmap, pv->pv_va)) != pa)
			panic("pmap_remove_all: bad mapping");
#endif
		pmap_remove(pv->pv_pmap, pv->pv_va, pv->pv_va + PAGE_SIZE);
	}
	splx(s);
}

/*
 *	Routine:	pmap_copy_on_write
 *	Function:
 *		Remove write privileges from all
 *		physical maps for this physical page.
 */
void
pmap_copy_on_write(pa)
	vm_offset_t pa;
{
#ifdef DEBUG
	if (pmapdebug & (PDB_FOLLOW|PDB_PROTECT))
		printf("pmap_copy_on_write(%x)", pa);
#endif
	pmap_changebit(pa, PG_RO, TRUE);
}

/*
 *	Set the physical protection on the
 *	specified range of this map as requested.
 */
void
pmap_protect(pmap, sva, eva, prot)
	register pmap_t	pmap;
	vm_offset_t	sva, eva;
	vm_prot_t	prot;
{
	register pt_entry_t *pte;
	register vm_offset_t va;
	register int ix;
	int i386prot;
	boolean_t firstpage = TRUE;

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

	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
		pmap_remove(pmap, sva, eva);
		return;
	}
	if (prot & VM_PROT_WRITE)
		return;

	for (va = sva; va < eva; va += PAGE_SIZE) {
		/*
		 * Page table page is not allocated.
		 * Skip it, we don't want to force allocation
		 * of unnecessary PTE pages just to set the protection.
		 */
		if (!pmap_pde_v(pmap_pde(pmap, va))) {
			/* XXX: avoid address wrap around */
			if (va >= i386_trunc_pdr((vm_offset_t)-1))
				break;
			va = i386_round_pdr(va + PAGE_SIZE) - PAGE_SIZE;
			continue;
		} else	pte = pmap_pte(pmap, va);

		/*
		 * Page not valid.  Again, skip it.
		 * Should we do this?  Or set protection anyway?
		 */
		if (!pmap_pte_v(pte))
			continue;

		ix = 0;
		i386prot = pte_prot(pmap, prot);
		if(va < UPT_MAX_ADDRESS)
			i386prot |= 2 /*PG_u*/;
		do {
			/* clear VAC here if PG_RO? */
			pmap_pte_set_prot(pte++, i386prot);
			/*TBIS(va + ix * I386_PAGE_SIZE);*/
		} while (++ix != i386pagesperpage);
	}
out:
	if (pmap == &curproc->p_vmspace->vm_pmap)
		pmap_activate(pmap, (struct pcb *)curproc->p_addr);
}

/*
 *	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, ix;
	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)",
		       pmap, va, pa, prot, wired);
	if(!pmap_isvalidphys(pa)) panic("invalid phys");
#endif
	if (pmap == NULL)
		return;

	if(va > VM_MAX_KERNEL_ADDRESS)panic("pmap_enter: toobig");
	/* also, should not muck with PTD va! */

#ifdef DEBUG
	if (pmap == kernel_pmap)
		enter_stats.kernel++;
	else
		enter_stats.user++;
#endif

	/*
	 * Page Directory table entry not valid, we need a new PT page
	 */
	if (!pmap_pde_v(pmap_pde(pmap, va))) {
		pg("ptdi %x", pmap->pm_pdir[PTDPTDI]);
	}

	pte = pmap_pte(pmap, va);
	opa = pmap_pte_pa(pte);
#ifdef DEBUG
	if (pmapdebug & PDB_ENTER)
		printf("enter: pte %x, *pte %x ", pte, *(int *)pte);
#endif

	/*
	 * Mapping has not changed, must be protection or wiring change.
	 */
	if (opa == pa) {
#ifdef DEBUG
		enter_stats.pwchange++;
#endif
		/*
		 * Wiring change, just update stats.
		 * We don't worry about wiring PT pages as they remain
		 * resident as long as there are valid mappings in them.
		 * Hence, if a user page is wired, the PT page will be also.
		 */
		if (wired && !pmap_pte_w(pte) || !wired && pmap_pte_w(pte)) {
#ifdef DEBUG
			if (pmapdebug & PDB_ENTER)
				pg("enter: wiring change -> %x ", wired);
#endif
			if (wired)
				pmap->pm_stats.wired_count++;
			else
				pmap->pm_stats.wired_count--;
#ifdef DEBUG
			enter_stats.wchange++;
#endif
		}
		goto validate;
	}

	/*
	 * Mapping has changed, invalidate old range and fall through to
	 * handle validating new mapping.
	 */
	if (opa) {
#ifdef DEBUG
		if (pmapdebug & PDB_ENTER)
			printf("enter: removing old mapping %x pa %x ", va, opa);
#endif
		pmap_remove(pmap, va, va + PAGE_SIZE);
#ifdef DEBUG
		enter_stats.mchange++;
#endif
	}

	/*
	 * Enter on the PV list if part of our managed memory
	 * Note that we raise IPL while manipulating pv_table
	 * since pmap_enter can be called at interrupt time.
	 */
	if (pa >= vm_first_phys && pa < vm_last_phys) {
		register pv_entry_t pv, npv;
		int s;

#ifdef DEBUG
		enter_stats.managed++;
#endif
		pv = pa_to_pvh(pa);
		s = splimp();
#ifdef DEBUG
		if (pmapdebug & PDB_ENTER)
			printf("enter: pv at %x: %x/%x/%x ",
			       pv, pv->pv_va, pv->pv_pmap, pv->pv_next);
#endif
		/*
		 * No entries yet, use header as the first entry
		 */
		if (pv->pv_pmap == NULL) {
#ifdef DEBUG
			enter_stats.firstpv++;
#endif
			pv->pv_va = va;
			pv->pv_pmap = pmap;
			pv->pv_next = NULL;
			pv->pv_flags = 0;
		}
		/*
		 * There is at least one other VA mapping this page.
		 * Place this entry after the header.
		 */
		else {
			/*printf("second time: ");*/
#ifdef DEBUG
			for (npv = pv; npv; npv = npv->pv_next)
				if (pmap == npv->pv_pmap && va == npv->pv_va)
					panic("pmap_enter: already in pv_tab");
#endif
			npv = (pv_entry_t)
				malloc(sizeof *npv, M_VMPVENT, M_NOWAIT);
			npv->pv_va = va;
			npv->pv_pmap = pmap;
			npv->pv_next = pv->pv_next;
			pv->pv_next = npv;
#ifdef DEBUG
			if (!npv->pv_next)
				enter_stats.secondpv++;
#endif
		splx(s);
		}
	}
	/*
	 * Assumption: if it is not part of our managed memory
	 * then it must be device memory which may be volitile.
	 */
	if (pmap_initialized) {
		checkpv = cacheable = FALSE;
#ifdef DEBUG
		enter_stats.unmanaged++;
#endif
	}

	/*
	 * Increment counters
	 */
	pmap->pm_stats.resident_count++;
	if (wired)
		pmap->pm_stats.wired_count++;

validate:
	/*
	 * Now validate mapping with desired protection/wiring.
	 * Assume uniform modified and referenced status for all
	 * I386 pages in a MACH page.
	 */
	npte = (pa & PG_FRAME) | pte_prot(pmap, prot) | PG_V;
	npte |= (*(int *)pte & (PG_M|PG_U));
	if (wired)
		npte |= PG_W;
	if(va < UPT_MIN_ADDRESS)
		npte |= PG_u;
	else if(va < UPT_MAX_ADDRESS)
		npte |= PG_u | PG_RW;
#ifdef DEBUG
	if (pmapdebug & PDB_ENTER)
		printf("enter: new pte value %x ", npte);
#endif
	ix = 0;
	do {
		*(int *)pte++ = npte;
		/*TBIS(va);*/
		npte += I386_PAGE_SIZE;
		va += I386_PAGE_SIZE;
	} while (++ix != i386pagesperpage);
	pte--;
#ifdef DEBUGx
cache, tlb flushes
#endif
/*pads(pmap);*/
	/*load_cr3(((struct pcb *)curproc->p_addr)->pcb_ptd);*/
	tlbflush();
}

/*
 *      pmap_page_protect:
 *
 *      Lower the permission for all mappings to a given page.
 */
void
pmap_page_protect(phys, prot)
        vm_offset_t     phys;
        vm_prot_t       prot;
{
        switch (prot) {
        case VM_PROT_READ:
        case VM_PROT_READ|VM_PROT_EXECUTE:
                pmap_copy_on_write(phys);
                break;
        case VM_PROT_ALL:
                break;
        default:
                pmap_remove_all(phys);
                break;
        }
}

/*
 *	Routine:	pmap_change_wiring
 *	Function:	Change the wiring attribute for a map/virtual-address
 *			pair.
 *	In/out conditions:
 *			The mapping must already exist in the pmap.
 */
void
pmap_change_wiring(pmap, va, wired)
	register pmap_t	pmap;
	vm_offset_t	va;
	boolean_t	wired;
{
	register pt_entry_t *pte;
	register int ix;

#ifdef DEBUG
	if (pmapdebug & PDB_FOLLOW)
		printf("pmap_change_wiring(%x, %x, %x)", pmap, va, wired);
#endif
	if (pmap == NULL)
		return;

	pte = pmap_pte(pmap, va);
#ifdef DEBUG
	/*
	 * Page table page is not allocated.
	 * Should this ever happen?  Ignore it for now,
	 * we don't want to force allocation of unnecessary PTE pages.
	 */
	if (!pmap_pde_v(pmap_pde(pmap, va))) {
		if (pmapdebug & PDB_PARANOIA)
			pg("pmap_change_wiring: invalid PDE for %x ", va);
		return;
	}
	/*
	 * Page not valid.  Should this ever happen?
	 * Just continue and change wiring anyway.
	 */
	if (!pmap_pte_v(pte)) {
		if (pmapdebug & PDB_PARANOIA)
			pg("pmap_change_wiring: invalid PTE for %x ", va);
	}
#endif
	if (wired && !pmap_pte_w(pte) || !wired && pmap_pte_w(pte)) {
		if (wired)
			pmap->pm_stats.wired_count++;
		else
			pmap->pm_stats.wired_count--;
	}
	/*
	 * Wiring is not a hardware characteristic so there is no need
	 * to invalidate TLB.
	 */
	ix = 0;
	do {
		pmap_pte_set_w(pte++, wired);
	} while (++ix != i386pagesperpage);
}

/*
 *	Routine:	pmap_pte
 *	Function:
 *		Extract the page table entry associated
 *		with the given map/virtual_address pair.
 * [ what about induced faults -wfj]
 */