memory.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

/*
 * Size memory and create the kernel page-tables on the fly while doing so.
 * Called from main(), this code should only be run by the bootstrap processor.
 */
#include "u.h"
#include "../port/lib.h"
#include "mem.h"
#include "dat.h"
#include "fns.h"
#include "io.h"
#include "ureg.h"

#define MEMDEBUG	0

enum {
	MemUPA		= 0,		/* unbacked physical address */
	MemRAM		= 1,		/* physical memory */
	MemUMB		= 2,		/* upper memory block (<16MB) */
	MemReserved	= 3,
	NMemType	= 4,

	KB		= 1024,

	MemMinMB	= 4,		/* minimum physical memory (<=4MB) */
	MemMaxMB	= 3*1024+768,	/* maximum physical memory to check */

	NMemBase	= 10,
};

typedef struct Map Map;
struct Map {
	ulong	size;
	ulong	addr;
};

typedef struct RMap RMap;
struct RMap {
	char*	name;
	Map*	map;
	Map*	mapend;

	Lock;
};

/* 
 * Memory allocation tracking.
 */
static Map mapupa[16];
static RMap rmapupa = {
	"unallocated unbacked physical memory",
	mapupa,
	&mapupa[nelem(mapupa)-1],
};

static Map xmapupa[16];
static RMap xrmapupa = {
	"unbacked physical memory",
	xmapupa,
	&xmapupa[nelem(xmapupa)-1],
};

static Map mapram[16];
static RMap rmapram = {
	"physical memory",
	mapram,
	&mapram[nelem(mapram)-1],
};

static Map mapumb[64];
static RMap rmapumb = {
	"upper memory block",
	mapumb,
	&mapumb[nelem(mapumb)-1],
};

static Map mapumbrw[16];
static RMap rmapumbrw = {
	"UMB device memory",
	mapumbrw,
	&mapumbrw[nelem(mapumbrw)-1],
};

void
mapprint(RMap *rmap)
{
	Map *mp;

	print("%s\n", rmap->name);	
	for(mp = rmap->map; mp->size; mp++)
		print("\t%8.8luX %8.8luX (%lud)\n", mp->addr, mp->addr+mp->size, mp->size);
}


void
memdebug(void)
{
	ulong maxpa, maxpa1, maxpa2;

	maxpa = (nvramread(0x18)<<8)|nvramread(0x17);
	maxpa1 = (nvramread(0x31)<<8)|nvramread(0x30);
	maxpa2 = (nvramread(0x16)<<8)|nvramread(0x15);
	print("maxpa = %luX -> %luX, maxpa1 = %luX maxpa2 = %luX\n",
		maxpa, MB+maxpa*KB, maxpa1, maxpa2);

	mapprint(&rmapram);
	mapprint(&rmapumb);
	mapprint(&rmapumbrw);
	mapprint(&rmapupa);
}

void
mapfree(RMap* rmap, ulong addr, ulong size)
{
	Map *mp;
	ulong t;

	if(size <= 0)
		return;

	lock(rmap);
	for(mp = rmap->map; mp->addr <= addr && mp->size; mp++)
		;

	if(mp > rmap->map && (mp-1)->addr+(mp-1)->size == addr){
		(mp-1)->size += size;
		if(addr+size == mp->addr){
			(mp-1)->size += mp->size;
			while(mp->size){
				mp++;
				(mp-1)->addr = mp->addr;
				(mp-1)->size = mp->size;
			}
		}
	}
	else{
		if(addr+size == mp->addr && mp->size){
			mp->addr -= size;
			mp->size += size;
		}
		else do{
			if(mp >= rmap->mapend){
				print("mapfree: %s: losing 0x%luX, %ld\n",
					rmap->name, addr, size);
				break;
			}
			t = mp->addr;
			mp->addr = addr;
			addr = t;
			t = mp->size;
			mp->size = size;
			mp++;
		}while(size = t);
	}
	unlock(rmap);
}

ulong
mapalloc(RMap* rmap, ulong addr, int size, int align)
{
	Map *mp;
	ulong maddr, oaddr;

	lock(rmap);
	for(mp = rmap->map; mp->size; mp++){
		maddr = mp->addr;

		if(addr){
			/*
			 * A specific address range has been given:
			 *   if the current map entry is greater then
			 *   the address is not in the map;
			 *   if the current map entry does not overlap
			 *   the beginning of the requested range then
			 *   continue on to the next map entry;
			 *   if the current map entry does not entirely
			 *   contain the requested range then the range
			 *   is not in the map.
			 */
			if(maddr > addr)
				break;
			if(mp->size < addr - maddr)	/* maddr+mp->size < addr, but no overflow */
				continue;
			if(addr - maddr > mp->size - size)	/* addr+size > maddr+mp->size, but no overflow */
				break;
			maddr = addr;
		}

		if(align > 0)
			maddr = ((maddr+align-1)/align)*align;
		if(mp->addr+mp->size-maddr < size)
			continue;

		oaddr = mp->addr;
		mp->addr = maddr+size;
		mp->size -= maddr-oaddr+size;
		if(mp->size == 0){
			do{
				mp++;
				(mp-1)->addr = mp->addr;
			}while((mp-1)->size = mp->size);
		}

		unlock(rmap);
		if(oaddr != maddr)
			mapfree(rmap, oaddr, maddr-oaddr);

		return maddr;
	}
	unlock(rmap);

	return 0;
}

/*
 * Allocate from the ram map directly to make page tables.
 * Called by mmuwalk during e820scan.
 */
void*
rampage(void)
{
	ulong m;
	
	m = mapalloc(&rmapram, 0, BY2PG, BY2PG);
	if(m == 0)
		return nil;
	return KADDR(m);
}

static void
umbexclude(void)
{
	int size;
	ulong addr;
	char *op, *p, *rptr;

	if((p = getconf("umbexclude")) == nil)
		return;

	while(p && *p != '\0' && *p != '\n'){
		op = p;
		addr = strtoul(p, &rptr, 0);
		if(rptr == nil || rptr == p || *rptr != '-'){
			print("umbexclude: invalid argument <%s>\n", op);
			break;
		}
		p = rptr+1;

		size = strtoul(p, &rptr, 0) - addr + 1;
		if(size <= 0){
			print("umbexclude: bad range <%s>\n", op);
			break;
		}
		if(rptr != nil && *rptr == ',')
			*rptr++ = '\0';
		p = rptr;

		mapalloc(&rmapumb, addr, size, 0);
	}
}

static void
umbscan(void)
{
	uchar *p;

	/*
	 * Scan the Upper Memory Blocks (0xA0000->0xF0000) for pieces
	 * which aren't used; they can be used later for devices which
	 * want to allocate some virtual address space.
	 * Check for two things:
	 * 1) device BIOS ROM. This should start with a two-byte header
	 *    of 0x55 0xAA, followed by a byte giving the size of the ROM
	 *    in 512-byte chunks. These ROM's must start on a 2KB boundary.
	 * 2) device memory. This is read-write.
	 * There are some assumptions: there's VGA memory at 0xA0000 and
	 * the VGA BIOS ROM is at 0xC0000. Also, if there's no ROM signature
	 * at 0xE0000 then the whole 64KB up to 0xF0000 is theoretically up
	 * for grabs; check anyway.
	 */
	p = KADDR(0xD0000);
	while(p < (uchar*)KADDR(0xE0000)){
		/*
		 * Test for 0x55 0xAA before poking obtrusively,
		 * some machines (e.g. Thinkpad X20) seem to map
		 * something dynamic here (cardbus?) causing weird
		 * problems if it is changed.
		 */
		if(p[0] == 0x55 && p[1] == 0xAA){
			p += p[2]*512;
			continue;
		}

		p[0] = 0xCC;
		p[2*KB-1] = 0xCC;
		if(p[0] != 0xCC || p[2*KB-1] != 0xCC){
			p[0] = 0x55;
			p[1] = 0xAA;
			p[2] = 4;
			if(p[0] == 0x55 && p[1] == 0xAA){
				p += p[2]*512;
				continue;
			}
			if(p[0] == 0xFF && p[1] == 0xFF)
				mapfree(&rmapumb, PADDR(p), 2*KB);
		}
		else
			mapfree(&rmapumbrw, PADDR(p), 2*KB);
		p += 2*KB;
	}

	p = KADDR(0xE0000);
	if(p[0] != 0x55 || p[1] != 0xAA){
		p[0] = 0xCC;
		p[64*KB-1] = 0xCC;
		if(p[0] != 0xCC && p[64*KB-1] != 0xCC)
			mapfree(&rmapumb, PADDR(p), 64*KB);
	}

	umbexclude();
}

static void
lowraminit(void)
{
	ulong n, pa, x;
	uchar *bda;

	/*
	 * Initialise the memory bank information for conventional memory
	 * (i.e. less than 640KB). The base is the first location after the
	 * bootstrap processor MMU information and the limit is obtained from
	 * the BIOS data area.
	 */
	x = PADDR(CPU0MACH+BY2PG);
	bda = (uchar*)KADDR(0x400);
	n = ((bda[0x14]<<8)|bda[0x13])*KB-x;
	mapfree(&rmapram, x, n);
	memset(KADDR(x), 0, n);			/* keep us honest */

	x = PADDR(PGROUND((ulong)end));
	pa = MemMinMB*MB;
	mapfree(&rmapram, x, pa-x);
	memset(KADDR(x), 0, pa-x);		/* keep us honest */
}

static void
ramscan(ulong maxmem)
{
	ulong *k0, kzero, map, maxkpa, maxpa, pa, *pte, *table, *va, vbase, x;
	int nvalid[NMemType];

	/*
	 * The bootstrap code has has created a prototype page
	 * table which maps the first MemMinMB of physical memory to KZERO.
	 * The page directory is at m->pdb and the first page of
	 * free memory is after the per-processor MMU information.
	 */
	pa = MemMinMB*MB;

	/*
	 * Check if the extended memory size can be obtained from the CMOS.
	 * If it's 0 then it's either not known or >= 64MB. Always check
	 * at least 24MB in case there's a memory gap (up to 8MB) below 16MB;
	 * in this case the memory from the gap is remapped to the top of
	 * memory.
	 * The value in CMOS is supposed to be the number of KB above 1MB.
	 */
	if(maxmem == 0){
		x = (nvramread(0x18)<<8)|nvramread(0x17);
		if(x == 0 || x >= (63*KB))
			maxpa = MemMaxMB*MB;
		else
			maxpa = MB+x*KB;
		if(maxpa < 24*MB)
			maxpa = 24*MB;
	}else
		maxpa = maxmem;
	maxkpa = (u32int)-KZERO;	/* 2^32 - KZERO */

	/*
	 * March up memory from MemMinMB to maxpa 1MB at a time,
	 * mapping the first page and checking the page can
	 * be written and read correctly. The page tables are created here
	 * on the fly, allocating from low memory as necessary.
	 */
	k0 = (ulong*)KADDR(0);
	kzero = *k0;
	map = 0;
	x = 0x12345678;
	memset(nvalid, 0, sizeof(nvalid));
	
	/*
	 * Can't map memory to KADDR(pa) when we're walking because
	 * can only use KADDR for relatively low addresses.  Instead,
	 * map each 4MB we scan to the virtual address range 4MB-8MB
	 * while we are scanning.
	 */
	vbase = 4*MB;
	while(pa < maxpa){
		/*
		 * Map the page. Use mapalloc(&rmapram, ...) to make
		 * the page table if necessary, it will be returned to the
		 * pool later if it isn't needed.  Map in a fixed range (the second 4M)
		 * because high physical addresses cannot be passed to KADDR.
		 */
		va = (void*)(vbase + pa%(4*MB));
		table = &m->pdb[PDX(va)];
		if(pa%(4*MB) == 0){
			if(map == 0 && (map = mapalloc(&rmapram, 0, BY2PG, BY2PG)) == 0)
				break;
			memset(KADDR(map), 0, BY2PG);
			*table = map|PTEWRITE|PTEVALID;
			memset(nvalid, 0, sizeof(nvalid));
		}
		table = KADDR(PPN(*table));
		pte = &table[PTX(va)];

		*pte = pa|PTEWRITE|PTEUNCACHED|PTEVALID;
		mmuflushtlb(PADDR(m->pdb));
		/*
		 * Write a pattern to the page and write a different
		 * pattern to a possible mirror at KZERO. If the data
		 * reads back correctly the chunk is some type of RAM (possibly
		 * a linearly-mapped VGA framebuffer, for instance...) and
		 * can be cleared and added to the memory pool. If not, the
		 * chunk is marked uncached and added to the UMB pool if <16MB
		 * or is marked invalid and added to the UPA pool.
		 */
		*va = x;
		*k0 = ~x;
		if(*va == x){
			nvalid[MemRAM] += MB/BY2PG;
			mapfree(&rmapram, pa, MB);

			do{
				*pte++ = pa|PTEWRITE|PTEVALID;
				pa += BY2PG;
			}while(pa % MB);
			mmuflushtlb(PADDR(m->pdb));
			/* memset(va, 0, MB); so damn slow to memset all of memory */
		}
		else if(pa < 16*MB){
			nvalid[MemUMB] += MB/BY2PG;
			mapfree(&rmapumb, pa, MB);

			do{
				*pte++ = pa|PTEWRITE|PTEUNCACHED|PTEVALID;
				pa += BY2PG;