libvhd.c

xen source 推出最新的VHD操作工具VHD-UTIL 实现源码,超强

		}
	}

	TEST_FAIL_AT(FAIL_REPARENT_LOCATOR);

	err = vhd_write_header(child, &child->header);
	if (err) {
		VHDLOG("error writing header for %s: %d\n", child->file, err);
		goto out;
	}

	err = 0;

out:
	free(ppath);
	return err;
}

static int
vhd_create_batmap(vhd_context_t *ctx)
{
	off64_t off;
	int err, map_bytes;
	vhd_batmap_header_t *header;

	if (!vhd_type_dynamic(ctx))
		return -EINVAL;

	map_bytes = (ctx->header.max_bat_size + 7) >> 3;
	header    = &ctx->batmap.header;

	memset(header, 0, sizeof(vhd_batmap_header_t));
	memcpy(header->cookie, VHD_BATMAP_COOKIE, sizeof(header->cookie));

	err = vhd_batmap_header_offset(ctx, &off);
	if (err)
		return err;

	header->batmap_offset  = off +
		vhd_bytes_padded(sizeof(vhd_batmap_header_t));
	header->batmap_size    = secs_round_up_no_zero(map_bytes);
	header->batmap_version = VHD_BATMAP_CURRENT_VERSION;

	map_bytes = vhd_sectors_to_bytes(header->batmap_size);

	err = posix_memalign((void **)&ctx->batmap.map,
			     VHD_SECTOR_SIZE, map_bytes);
	if (err) {
		ctx->batmap.map = NULL;
		return -err;
	}

	memset(ctx->batmap.map, 0, map_bytes);

	return vhd_write_batmap(ctx, &ctx->batmap);
}

static int
vhd_create_bat(vhd_context_t *ctx)
{
	int i, err;
	size_t size;

	if (!vhd_type_dynamic(ctx))
		return -EINVAL;

	size = vhd_bytes_padded(ctx->header.max_bat_size * sizeof(uint32_t));
	err  = posix_memalign((void **)&ctx->bat.bat, VHD_SECTOR_SIZE, size);
	if (err) {
		ctx->bat.bat = NULL;
		return err;
	}

	memset(ctx->bat.bat, 0, size);
	for (i = 0; i < ctx->header.max_bat_size; i++)
		ctx->bat.bat[i] = DD_BLK_UNUSED;

	err = vhd_seek(ctx, ctx->header.table_offset, SEEK_SET);
	if (err)
		return err;

	ctx->bat.entries = ctx->header.max_bat_size;
	ctx->bat.spb     = ctx->header.block_size >> VHD_SECTOR_SHIFT;

	return vhd_write_bat(ctx, &ctx->bat);
}

static int
vhd_initialize_fixed_disk(vhd_context_t *ctx)
{
	char *buf;
	int i, err;

	if (ctx->footer.type != HD_TYPE_FIXED)
		return -EINVAL;

	err = vhd_seek(ctx, 0, SEEK_SET);
	if (err)
		return err;

	buf = mmap(0, VHD_BLOCK_SIZE, PROT_READ,
		   MAP_SHARED | MAP_ANONYMOUS, -1, 0);
	if (buf == MAP_FAILED)
		return -errno;

	for (i = 0; i < ctx->footer.curr_size >> VHD_BLOCK_SHIFT; i++) {
		err = vhd_write(ctx, buf, VHD_BLOCK_SIZE);
		if (err)
			goto out;
	}

	err = 0;

out:
	munmap(buf, VHD_BLOCK_SIZE);
	return err;
}

int 
vhd_get_phys_size(vhd_context_t *ctx, off64_t *size)
{
	int err;

	if ((err = vhd_end_of_data(ctx, size)))
		return err;
	*size += sizeof(vhd_footer_t);
	return 0;
}

int 
vhd_set_phys_size(vhd_context_t *ctx, off64_t size)
{
	off64_t phys_size;
	int err;

	err = vhd_get_phys_size(ctx, &phys_size);
	if (err)
		return err;
	if (size < phys_size) {
		// would result in data loss
		VHDLOG("ERROR: new size (%"PRIu64") < phys size (%"PRIu64")\n",
				size, phys_size);
		return -EINVAL;
	}
	return vhd_write_footer_at(ctx, &ctx->footer, 
			size - sizeof(vhd_footer_t));
}

static int
__vhd_create(const char *name, const char *parent, uint64_t bytes, int type,
		vhd_flag_creat_t flags)
{
	int err;
	off64_t off;
	vhd_context_t ctx;
	vhd_footer_t *footer;
	vhd_header_t *header;
	uint64_t size, blks;

	switch (type) {
	case HD_TYPE_DIFF:
		if (!parent)
			return -EINVAL;
	case HD_TYPE_FIXED:
	case HD_TYPE_DYNAMIC:
		break;
	default:
		return -EINVAL;
	}

	if (strnlen(name, VHD_MAX_NAME_LEN - 1) == VHD_MAX_NAME_LEN - 1)
		return -ENAMETOOLONG;

	memset(&ctx, 0, sizeof(vhd_context_t));
	footer = &ctx.footer;
	header = &ctx.header;
	blks   = (bytes + VHD_BLOCK_SIZE - 1) >> VHD_BLOCK_SHIFT;
	size   = blks << VHD_BLOCK_SHIFT;

	ctx.fd = open(name, O_WRONLY | O_CREAT |
		      O_TRUNC | O_LARGEFILE | O_DIRECT, 0644);
	if (ctx.fd == -1)
		return -errno;

	ctx.file = strdup(name);
	if (!ctx.file) {
		err = -ENOMEM;
		goto out;
	}

	err = vhd_test_file_fixed(ctx.file, &ctx.is_block);
	if (err)
		goto out;

	vhd_initialize_footer(&ctx, type, size);

	if (type == HD_TYPE_FIXED) {
		err = vhd_initialize_fixed_disk(&ctx);
		if (err)
			goto out;
	} else {
		int raw = vhd_flag_test(flags, VHD_FLAG_CREAT_PARENT_RAW);
		err = vhd_initialize_header(&ctx, parent, size, raw);
		if (err)
			goto out;

		err = vhd_write_footer_at(&ctx, &ctx.footer, 0);
		if (err)
			goto out;

		err = vhd_write_header_at(&ctx, &ctx.header, VHD_SECTOR_SIZE);
		if (err)
			goto out;

		err = vhd_create_batmap(&ctx);
		if (err)
			goto out;

		err = vhd_create_bat(&ctx);
		if (err)
			goto out;

		if (type == HD_TYPE_DIFF) {
			err = vhd_write_parent_locators(&ctx, parent);
			if (err)
				goto out;
		}

		/* write header again since it may have changed */
		err = vhd_write_header_at(&ctx, &ctx.header, VHD_SECTOR_SIZE);
		if (err)
			goto out;
	}

	err = vhd_seek(&ctx, 0, SEEK_END);
	if (err)
		goto out;

	off = vhd_position(&ctx);
	if (off == (off64_t)-1) {
		err = -errno;
		goto out;
	}

	if (ctx.is_block)
		off -= sizeof(vhd_footer_t);

	err = vhd_write_footer_at(&ctx, &ctx.footer, off);
	if (err)
		goto out;

	err = 0;

out:
	vhd_close(&ctx);
	if (err && !ctx.is_block)
		unlink(name);
	return err;
}

int
vhd_create(const char *name, uint64_t bytes, int type, vhd_flag_creat_t flags)
{
	return __vhd_create(name, NULL, bytes, type, flags);
}

int
vhd_snapshot(const char *name, uint64_t bytes, const char *parent,
		vhd_flag_creat_t flags)
{
	return __vhd_create(name, parent, bytes, HD_TYPE_DIFF, flags);
}

static int
__vhd_io_fixed_read(vhd_context_t *ctx,
		    char *buf, uint64_t sec, uint32_t secs)
{
	int err;

	err = vhd_seek(ctx, vhd_sectors_to_bytes(sec), SEEK_SET);
	if (err)
		return err;

	return vhd_read(ctx, buf, vhd_sectors_to_bytes(secs));
}

static void
__vhd_io_dynamic_copy_data(vhd_context_t *ctx,
			   char *map, int map_off,
			   char *bitmap, int bitmap_off,
			   char *dst, char *src, int secs)
{
	int i;

	for (i = 0; i < secs; i++) {
		if (test_bit(map, map_off + i))
			goto next;

		if (ctx && !vhd_bitmap_test(ctx, bitmap, bitmap_off + i))
			goto next;

		memcpy(dst, src, VHD_SECTOR_SIZE);
		set_bit(map, map_off + i);

	next:
		src += VHD_SECTOR_SIZE;
		dst += VHD_SECTOR_SIZE;
	}
}

static int
__vhd_io_dynamic_read_link(vhd_context_t *ctx, char *map,
			   char *buf, uint64_t sector, uint32_t secs)
{
	off64_t off;
	uint32_t blk, sec;
	int err, cnt, map_off;
	char *bitmap, *data, *src;

	map_off = 0;

	do {
		blk    = sector / ctx->spb;
		sec    = sector % ctx->spb;
		off    = ctx->bat.bat[blk];
		data   = NULL;
		bitmap = NULL;

		if (off == DD_BLK_UNUSED) {
			cnt = MIN(secs, ctx->spb);
			goto next;
		}

		err = vhd_read_bitmap(ctx, blk, &bitmap);
		if (err)
			return err;

		err = vhd_read_block(ctx, blk, &data);
		if (err) {
			free(bitmap);
			return err;
		}

		cnt = MIN(secs, ctx->spb - sec);
		src = data + vhd_sectors_to_bytes(sec);

		__vhd_io_dynamic_copy_data(ctx,
					   map, map_off,
					   bitmap, sec,
					   buf, src, cnt);

	next:
		free(data);
		free(bitmap);

		secs    -= cnt;
		sector  += cnt;
		map_off += cnt;
		buf     += vhd_sectors_to_bytes(cnt);

	} while (secs);

	return 0;
}

static int
__raw_read_link(char *filename,
		char *map, char *buf, uint64_t sec, uint32_t secs)
{
	int fd, err;
	off64_t off;
	uint64_t size;
	char *data;

	err = 0;
	errno = 0;
	fd = open(filename, O_RDONLY | O_DIRECT | O_LARGEFILE);
	if (fd == -1) {
		VHDLOG("%s: failed to open: %d\n", filename, -errno);
		return -errno;
	}

	off = lseek64(fd, vhd_sectors_to_bytes(sec), SEEK_SET);
	if (off == (off64_t)-1) {
		VHDLOG("%s: seek(0x%08"PRIx64") failed: %d\n",
		       filename, vhd_sectors_to_bytes(sec), -errno);
		err = -errno;
		goto close;
	}

	size = vhd_sectors_to_bytes(secs);
	err = posix_memalign((void **)&data, VHD_SECTOR_SIZE, size);
	if (err)
		goto close;

	err = read(fd, data, size);
	if (err != size) {
		VHDLOG("%s: reading of %"PRIu64" returned %d, errno: %d\n",
				filename, size, err, -errno);
		free(data);
		err = errno ? -errno : -EIO;
		goto close;
	}
	__vhd_io_dynamic_copy_data(NULL, map, 0, NULL, 0, buf, data, secs);
	free(data);
	err = 0;

close:
	close(fd);
	return err;
}

static int
__vhd_io_dynamic_read(vhd_context_t *ctx,
		      char *buf, uint64_t sec, uint32_t secs)
{
	int err;
	uint32_t i, done;
	char *map, *next;
	vhd_context_t parent, *vhd;

	err  = vhd_get_bat(ctx);
	if (err)
		return err;

	vhd  = ctx;
	next = NULL;
	map  = calloc(1, secs << (VHD_SECTOR_SHIFT - 3));
	if (!map)
		return -ENOMEM;

	memset(buf, 0, vhd_sectors_to_bytes(secs));

	for (;;) {
		err = __vhd_io_dynamic_read_link(vhd, map, buf, sec, secs);
		if (err)
			goto close;

		for (done = 0, i = 0; i < secs; i++)
			if (test_bit(map, i))
				done++;

		if (done == secs) {
			err = 0;
			goto close;
		}

		if (vhd->footer.type == HD_TYPE_DIFF) {
			err = vhd_parent_locator_get(vhd, &next);
			if (err)
				goto close;
			if (vhd_parent_raw(vhd)) {
				err = __raw_read_link(next, map, buf, sec,
						secs);
				goto close;
			}
		} else {
			err = 0;
			goto close;
		}

		if (vhd != ctx)
			vhd_close(vhd);
		vhd = &parent;

		err = vhd_open(vhd, next, VHD_OPEN_RDONLY);
		if (err)
			goto out;

		err = vhd_get_bat(vhd);
		if (err)
			goto close;

		free(next);
		next = NULL;
	}

close:
	if (vhd != ctx)
		vhd_close(vhd);
out:
	free(map);
	free(next);
	return err;
}

int
vhd_io_read(vhd_context_t *ctx, char *buf, uint64_t sec, uint32_t secs)
{
	if (vhd_sectors_to_bytes(sec + secs) > ctx->footer.curr_size)
		return -ERANGE;

	if (!vhd_type_dynamic(ctx))
		return __vhd_io_fixed_read(ctx, buf, sec, secs);

	return __vhd_io_dynamic_read(ctx, buf, sec, secs);
}

static int
__vhd_io_fixed_write(vhd_context_t *ctx,
		     char *buf, uint64_t sec, uint32_t secs)
{
	int err;

	err = vhd_seek(ctx, vhd_sectors_to_bytes(sec), SEEK_SET);
	if (err)
		return err;

	return vhd_write(ctx, buf, vhd_sectors_to_bytes(secs));
}

static int
__vhd_io_allocate_block(vhd_context_t *ctx, uint32_t block)
{
	char *buf;
	size_t size;
	off64_t off, max;
	int i, err, gap, spp;

	spp = getpagesize() >> VHD_SECTOR_SHIFT;

	err = vhd_end_of_data(ctx, &max);
	if (err)
		return err;

	gap   = 0;
	off   = max;
	max >>= VHD_SECTOR_SHIFT;

	/* data region of segment should begin on page boundary */
	if ((max + ctx->bm_secs) % spp) {
		gap  = (spp - ((max + ctx->bm_secs) % spp));
		max += gap;
	}

	err = vhd_seek(ctx, off, SEEK_SET);
	if (err)
		return err;

	size = vhd_sectors_to_bytes(ctx->spb + ctx->bm_secs + gap);
	buf  = mmap(0, size, PROT_READ, MAP_SHARED | MAP_ANONYMOUS, -1, 0);
	if (buf == MAP_FAILED)
		return -errno;

	err = vhd_write(ctx, buf, size);
	if (err)
		goto out;

	ctx->bat.bat[block] = max;
	err = vhd_write_bat(ctx, &ctx->bat);
	if (err)
		goto out;

	err = 0;

out:
	munmap(buf, size);
	return err;
}

static int
__vhd_io_dynamic_write(vhd_context_t *ctx,
		       char *buf, uint64_t sector, uint32_t secs)
{
	char *map;
	off64_t off;
	uint32_t blk, sec;
	int i, err, cnt, ret;

	if (vhd_sectors_to_bytes(sector + secs) > ctx->footer.curr_size)
		return -ERANGE;

	err = vhd_get_bat(ctx);
	if (err)
		return err;

	if (vhd_has_batmap(ctx)) {
		err = vhd_get_batmap(ctx);
		if (err)
			return err;
	}

	do {
		blk = sector / ctx->spb;
		sec = sector % ctx->spb;

		off = ctx->bat.bat[blk];
		if (off == DD_BLK_UNUSED) {
			err = __vhd_io_allocate_block(ctx, blk);
			if (err)
				return err;

			off = ctx->bat.bat[blk];
		}

		off += ctx->bm_secs + sec;
		err  = vhd_seek(ctx, vhd_sectors_to_bytes(off), SEEK_SET);
		if (err)
			return err;

		cnt = MIN(secs, ctx->spb - sec);
		err = vhd_write(ctx, buf, vhd_sectors_to_bytes(cnt));
		if (err)
			return err;

		if (vhd_has_batmap(ctx) &&
		    vhd_batmap_test(ctx, &ctx->batmap, blk))
			goto next;

		err = vhd_read_bitmap(ctx, blk, &map);
		if (err)
			return err;

		for (i = 0; i < cnt; i++)
			vhd_bitmap_set(ctx, map, sec + i);

		err = vhd_write_bitmap(ctx, blk, map);
		if (err)
			goto fail;

		if (vhd_has_batmap(ctx)) {
			for (i = 0; i < ctx->spb; i++)
				if (!vhd_bitmap_test(ctx, map, i)) {
					free(map);
					goto next;
				}

			vhd_batmap_set(ctx, &ctx->batmap, blk);
			err = vhd_write_batmap(ctx, &ctx->batmap);
			if (err)
				goto fail;
		}

		free(map);
		map = NULL;

	next:
		secs   -= cnt;
		sector += cnt;
		buf    += vhd_sectors_to_bytes(cnt);
	} while (secs);

	err = 0;

out:
	ret = vhd_write_footer(ctx, &ctx->footer);
	return (err ? err : ret);

fail:
	free(map);
	goto out;
}

int
vhd_io_write(vhd_context_t *ctx, char *buf, uint64_t sec, uint32_t secs)
{
	if (vhd_sectors_to_bytes(sec + secs) > ctx->footer.curr_size)
		return -ERANGE;

	if (!vhd_type_dynamic(ctx))
		return __vhd_io_fixed_write(ctx, buf, sec, secs);

	return __vhd_io_dynamic_write(ctx, buf, sec, secs);
}