mmu.c

这是一个同样来自贝尔实验室的和UNIX有着渊源的操作系统, 其简洁的设计和实现易于我们学习和理解

 * Device mappings are shared by all procs and processors and
 * live in the virtual range VMAP to VMAP+VMAPSIZE.  The master
 * copy of the mappings is stored in mach0->pdb, and they are
 * paged in from there as necessary by vmapsync during faults.
 */

static Lock vmaplock;

static int findhole(ulong *a, int n, int count);
static ulong vmapalloc(ulong size);
static void pdbunmap(ulong*, ulong, int);

/*
 * Add a device mapping to the vmap range.
 */
void*
vmap(ulong pa, int size)
{
	int osize;
	ulong o, va;
	
	/*
	 * might be asking for less than a page.
	 */
	osize = size;
	o = pa & (BY2PG-1);
	pa -= o;
	size += o;

	size = ROUND(size, BY2PG);
	if(pa == 0){
		print("vmap pa=0 pc=%#.8lux\n", getcallerpc(&pa));
		return nil;
	}
	ilock(&vmaplock);
	if((va = vmapalloc(size)) == 0 
	|| pdbmap(MACHP(0)->pdb, pa|PTEUNCACHED|PTEWRITE, va, size) < 0){
		iunlock(&vmaplock);
		return 0;
	}
	iunlock(&vmaplock);
	/* avoid trap on local processor
	for(i=0; i<size; i+=4*MB)
		vmapsync(va+i);
	*/
	USED(osize);
//	print("  vmap %#.8lux %d => %#.8lux\n", pa+o, osize, va+o);
	return (void*)(va + o);
}

static int
findhole(ulong *a, int n, int count)
{
	int have, i;
	
	have = 0;
	for(i=0; i<n; i++){
		if(a[i] == 0)
			have++;
		else
			have = 0;
		if(have >= count)
			return i+1 - have;
	}
	return -1;
}

/*
 * Look for free space in the vmap.
 */
static ulong
vmapalloc(ulong size)
{
	int i, n, o;
	ulong *vpdb;
	int vpdbsize;
	
	vpdb = &MACHP(0)->pdb[PDX(VMAP)];
	vpdbsize = VMAPSIZE/(4*MB);

	if(size >= 4*MB){
		n = (size+4*MB-1) / (4*MB);
		if((o = findhole(vpdb, vpdbsize, n)) != -1)
			return VMAP + o*4*MB;
		return 0;
	}
	n = (size+BY2PG-1) / BY2PG;
	for(i=0; i<vpdbsize; i++)
		if((vpdb[i]&PTEVALID) && !(vpdb[i]&PTESIZE))
			if((o = findhole(KADDR(PPN(vpdb[i])), WD2PG, n)) != -1)
				return VMAP + i*4*MB + o*BY2PG;
	if((o = findhole(vpdb, vpdbsize, 1)) != -1)
		return VMAP + o*4*MB;
		
	/*
	 * could span page directory entries, but not worth the trouble.
	 * not going to be very much contention.
	 */
	return 0;
}

/*
 * Remove a device mapping from the vmap range.
 * Since pdbunmap does not remove page tables, just entries,
 * the call need not be interlocked with vmap.
 */
void
vunmap(void *v, int size)
{
	int i;
	ulong va, o;
	Mach *nm;
	Proc *p;
	
	/*
	 * might not be aligned
	 */
	va = (ulong)v;
	o = va&(BY2PG-1);
	va -= o;
	size += o;
	size = ROUND(size, BY2PG);
	
	if(size < 0 || va < VMAP || va+size > VMAP+VMAPSIZE)
		panic("vunmap va=%#.8lux size=%#x pc=%#.8lux\n",
			va, size, getcallerpc(&va));

	pdbunmap(MACHP(0)->pdb, va, size);
	
	/*
	 * Flush mapping from all the tlbs and copied pdbs.
	 * This can be (and is) slow, since it is called only rarely.
	 */
	for(i=0; i<conf.nproc; i++){
		p = proctab(i);
		if(p->state == Dead)
			continue;
		if(p != up)
			p->newtlb = 1;
	}
	for(i=0; i<conf.nmach; i++){
		nm = MACHP(i);
		if(nm != m)
			nm->flushmmu = 1;
	}
	flushmmu();
	for(i=0; i<conf.nmach; i++){
		nm = MACHP(i);
		if(nm != m)
			while((active.machs&(1<<nm->machno)) && nm->flushmmu)
				;
	}
}

/*
 * Add kernel mappings for pa -> va for a section of size bytes.
 */
int
pdbmap(ulong *pdb, ulong pa, ulong va, int size)
{
	int pse;
	ulong pgsz, *pte, *table;
	ulong flag, off;
	
	flag = pa&0xFFF;
	pa &= ~0xFFF;

	if((MACHP(0)->cpuiddx & 0x08) && (getcr4() & 0x10))
		pse = 1;
	else
		pse = 0;

	for(off=0; off<size; off+=pgsz){
		table = &pdb[PDX(va+off)];
		if((*table&PTEVALID) && (*table&PTESIZE))
			panic("vmap: va=%#.8lux pa=%#.8lux pde=%#.8lux",
				va+off, pa+off, *table);

		/*
		 * Check if it can be mapped using a 4MB page:
		 * va, pa aligned and size >= 4MB and processor can do it.
		 */
		if(pse && (pa+off)%(4*MB) == 0 && (va+off)%(4*MB) == 0 && (size-off) >= 4*MB){
			*table = (pa+off)|flag|PTESIZE|PTEVALID;
			pgsz = 4*MB;
		}else{
			pte = mmuwalk(pdb, va+off, 2, 1);
			if(*pte&PTEVALID)
				panic("vmap: va=%#.8lux pa=%#.8lux pte=%#.8lux",
					va+off, pa+off, *pte);
			*pte = (pa+off)|flag|PTEVALID;
			pgsz = BY2PG;
		}
	}
	return 0;
}

/*
 * Remove mappings.  Must already exist, for sanity.
 * Only used for kernel mappings, so okay to use KADDR.
 */
static void
pdbunmap(ulong *pdb, ulong va, int size)
{
	ulong vae;
	ulong *table;
	
	vae = va+size;
	while(va < vae){
		table = &pdb[PDX(va)];
		if(!(*table & PTEVALID)){
			panic("vunmap: not mapped");
			/* 
			va = (va+4*MB-1) & ~(4*MB-1);
			continue;
			*/
		}
		if(*table & PTESIZE){
			*table = 0;
			va = (va+4*MB-1) & ~(4*MB-1);
			continue;
		}
		table = KADDR(PPN(*table));
		if(!(table[PTX(va)] & PTEVALID))
			panic("vunmap: not mapped");
		table[PTX(va)] = 0;
		va += BY2PG;
	}
}

/*
 * Handle a fault by bringing vmap up to date.
 * Only copy pdb entries and they never go away,
 * so no locking needed.
 */
int
vmapsync(ulong va)
{
	ulong entry, *table;

	if(va < VMAP || va >= VMAP+VMAPSIZE)
		return 0;

	entry = MACHP(0)->pdb[PDX(va)];
	if(!(entry&PTEVALID))
		return 0;
	if(!(entry&PTESIZE)){
		/* make sure entry will help the fault */
		table = KADDR(PPN(entry));
		if(!(table[PTX(va)]&PTEVALID))
			return 0;
	}
	vpd[PDX(va)] = entry;
	/*
	 * TLB doesn't cache negative results, so no flush needed.
	 */
	return 1;
}


/*
 * KMap is used to map individual pages into virtual memory.
 * It is rare to have more than a few KMaps at a time (in the 
 * absence of interrupts, only two at a time are ever used,
 * but interrupts can stack).  The mappings are local to a process,
 * so we can use the same range of virtual address space for
 * all processes without any coordination.
 */
#define kpt (vpt+VPTX(KMAP))
#define NKPT (KMAPSIZE/BY2PG)

KMap*
kmap(Page *page)
{
	int i, o, s;

	if(up == nil)
		panic("kmap: up=0 pc=%#.8lux", getcallerpc(&page));
	if(up->mmupdb == nil)
		upallocpdb();
	if(up->nkmap < 0)
		panic("kmap %lud %s: nkmap=%d", up->pid, up->text, up->nkmap);
	
	/*
	 * Splhi shouldn't be necessary here, but paranoia reigns.
	 * See comment in putmmu above.
	 */
	s = splhi();
	up->nkmap++;
	if(!(vpd[PDX(KMAP)]&PTEVALID)){
		/* allocate page directory */
		if(KMAPSIZE > BY2XPG)
			panic("bad kmapsize");
		if(up->kmaptable != nil)
			panic("kmaptable");
		spllo();
		up->kmaptable = newpage(0, 0, 0);
		splhi();
		vpd[PDX(KMAP)] = up->kmaptable->pa|PTEWRITE|PTEVALID;
		flushpg((ulong)kpt);
		memset(kpt, 0, BY2PG);
		kpt[0] = page->pa|PTEWRITE|PTEVALID;
		up->lastkmap = 0;
		splx(s);
		return (KMap*)KMAP;
	}
	if(up->kmaptable == nil)
		panic("no kmaptable");
	o = up->lastkmap+1;
	for(i=0; i<NKPT; i++){
		if(kpt[(i+o)%NKPT] == 0){
			o = (i+o)%NKPT;
			kpt[o] = page->pa|PTEWRITE|PTEVALID;
			up->lastkmap = o;
			splx(s);
			return (KMap*)(KMAP+o*BY2PG);
		}
	}
	panic("out of kmap");
	return nil;
}

void
kunmap(KMap *k)
{
	ulong va;

	va = (ulong)k;
	if(up->mmupdb == nil || !(vpd[PDX(KMAP)]&PTEVALID))
		panic("kunmap: no kmaps");
	if(va < KMAP || va >= KMAP+KMAPSIZE)
		panic("kunmap: bad address %#.8lux pc=%#.8lux", va, getcallerpc(&k));
	if(!(vpt[VPTX(va)]&PTEVALID))
		panic("kunmap: not mapped %#.8lux pc=%#.8lux", va, getcallerpc(&k));
	up->nkmap--;
	if(up->nkmap < 0)
		panic("kunmap %lud %s: nkmap=%d", up->pid, up->text, up->nkmap);
	vpt[VPTX(va)] = 0;
	flushpg(va);
}

/*
 * Temporary one-page mapping used to edit page directories.
 *
 * The fasttmp #define controls whether the code optimizes
 * the case where the page is already mapped in the physical
 * memory window.  
 */
#define fasttmp 1

void*
tmpmap(Page *p)
{
	ulong i;
	ulong *entry;
	
	if(islo())
		panic("tmpaddr: islo");

	if(fasttmp && p->pa < -KZERO)
		return KADDR(p->pa);

	/*
	 * PDX(TMPADDR) == PDX(MACHADDR), so this
	 * entry is private to the processor and shared 
	 * between up->mmupdb (if any) and m->pdb.
	 */
	entry = &vpt[VPTX(TMPADDR)];
	if(!(*entry&PTEVALID)){
		for(i=KZERO; i<=CPU0MACH; i+=BY2PG)
			print("%.8lux: *%.8lux=%.8lux (vpt=%.8lux index=%.8lux)\n", i, &vpt[VPTX(i)], vpt[VPTX(i)], vpt, VPTX(i));
		panic("tmpmap: no entry");
	}
	if(PPN(*entry) != PPN(TMPADDR-KZERO))
		panic("tmpmap: already mapped entry=%#.8lux", *entry);
	*entry = p->pa|PTEWRITE|PTEVALID;
	flushpg(TMPADDR);
	return (void*)TMPADDR;
}

void
tmpunmap(void *v)
{
	ulong *entry;
	
	if(islo())
		panic("tmpaddr: islo");
	if(fasttmp && (ulong)v >= KZERO && v != (void*)TMPADDR)
		return;
	if(v != (void*)TMPADDR)
		panic("tmpunmap: bad address");
	entry = &vpt[VPTX(TMPADDR)];
	if(!(*entry&PTEVALID) || PPN(*entry) == PPN(PADDR(TMPADDR)))
		panic("tmpmap: not mapped entry=%#.8lux", *entry);
	*entry = PPN(TMPADDR-KZERO)|PTEWRITE|PTEVALID;
	flushpg(TMPADDR);
}

/*
 * These could go back to being macros once the kernel is debugged,
 * but the extra checking is nice to have.
 */
void*
kaddr(ulong pa)
{
	if(pa > (ulong)-KZERO)
		panic("kaddr: pa=%#.8lux", pa);
	return (void*)(pa+KZERO);
}

ulong
paddr(void *v)
{
	ulong va;
	
	va = (ulong)v;
	if(va < KZERO)
		panic("paddr: va=%#.8lux pc=%#.8lux", va, getcallerpc(&va));
	return va-KZERO;
}

/*
 * More debugging.
 */
void
countpagerefs(ulong *ref, int print)
{
	int i, n;
	Mach *mm;
	Page *pg;
	Proc *p;
	
	n = 0;
	for(i=0; i<conf.nproc; i++){
		p = proctab(i);
		if(p->mmupdb){
			if(print){
				if(ref[pagenumber(p->mmupdb)])
					iprint("page %#.8lux is proc %d (pid %lud) pdb\n",
						p->mmupdb->pa, i, p->pid);
				continue;
			}
			if(ref[pagenumber(p->mmupdb)]++ == 0)
				n++;
			else
				iprint("page %#.8lux is proc %d (pid %lud) pdb but has other refs!\n",
					p->mmupdb->pa, i, p->pid);
		}
		if(p->kmaptable){
			if(print){
				if(ref[pagenumber(p->kmaptable)])
					iprint("page %#.8lux is proc %d (pid %lud) kmaptable\n",
						p->kmaptable->pa, i, p->pid);
				continue;
			}
			if(ref[pagenumber(p->kmaptable)]++ == 0)
				n++;
			else
				iprint("page %#.8lux is proc %d (pid %lud) kmaptable but has other refs!\n",
					p->kmaptable->pa, i, p->pid);
		}
		for(pg=p->mmuused; pg; pg=pg->next){
			if(print){
				if(ref[pagenumber(pg)])
					iprint("page %#.8lux is on proc %d (pid %lud) mmuused\n",
						pg->pa, i, p->pid);
				continue;
			}
			if(ref[pagenumber(pg)]++ == 0)
				n++;
			else
				iprint("page %#.8lux is on proc %d (pid %lud) mmuused but has other refs!\n",
					pg->pa, i, p->pid);
		}
		for(pg=p->mmufree; pg; pg=pg->next){
			if(print){
				if(ref[pagenumber(pg)])
					iprint("page %#.8lux is on proc %d (pid %lud) mmufree\n",
						pg->pa, i, p->pid);
				continue;
			}
			if(ref[pagenumber(pg)]++ == 0)
				n++;
			else
				iprint("page %#.8lux is on proc %d (pid %lud) mmufree but has other refs!\n",
					pg->pa, i, p->pid);
		}
	}
	if(!print)
		iprint("%d pages in proc mmu\n", n);
	n = 0;
	for(i=0; i<conf.nmach; i++){
		mm = MACHP(i);
		for(pg=mm->pdbpool; pg; pg=pg->next){
			if(print){
				if(ref[pagenumber(pg)])
					iprint("page %#.8lux is in cpu%d pdbpool\n",
						pg->pa, i);
				continue;
			}
			if(ref[pagenumber(pg)]++ == 0)
				n++;
			else
				iprint("page %#.8lux is in cpu%d pdbpool but has other refs!\n",
					pg->pa, i);
		}
	}
	if(!print){
		iprint("%d pages in mach pdbpools\n", n);
		for(i=0; i<conf.nmach; i++)
			iprint("cpu%d: %d pdballoc, %d pdbfree\n",
				i, MACHP(i)->pdballoc, MACHP(i)->pdbfree);
	}
}

void
checkfault(ulong, ulong)
{
}