Re: [PATCH 1/2] fs:regfs: add register easy filesystem

[zc <zoucao@linux.alibaba.com>]

在 2020/10/29 上午10:42, Al Viro 写道:
> On Tue, Oct 20, 2020 at 02:30:07PM +0800, Zou Cao wrote:
>> +ssize_t regfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
>> +{
>> +	struct file *file = iocb->ki_filp;
>> +	struct inode *inode = file->f_mapping->host;
>> +	ssize_t ret;
>> +
>> +	inode_lock(inode);
>> +	ret = generic_write_checks(iocb, from);
>> +	if (ret > 0)
>> +		ret = __generic_file_write_iter(iocb, from);
>> +	inode_unlock(inode);
>> +
>> +	if (ret > 0)
>> +		ret = generic_write_sync(iocb, ret);
>> +	return ret;
>> +}
> Huh?  How is that different from generic_file_write_iter()?  And who's
> using it, anyway?
my bug,  remove it.
>> +	struct regfs_inode_info  *info = REGFS_I(mapping->host);
>> +	char str[67];
>> +	unsigned long val = 0;
>> +	loff_t pos = *ppos;
>> +	size_t res;
>> +
>> +	if (pos < 0)
>> +		return -EINVAL;
>> +	if (pos >= len || len > 66)
>> +		return 0;
> This is completely bogus.  "If current position is greater than the
> length of string we are asking to write, quietly return 0"?
Yes, fixed it.
>> +	res = copy_from_user(str, buf, len);
>> +	if (res)
>> +		return -EFAULT;
>> +	str[len] = 0;
>> +
>> +	if (kstrtoul(str, 16, &val) < 0)
>> +		return -EINVAL;
> Where does 67 come from?  If you are expecting a hexadecimal representation
> of a unsigned long on arm64, you should have at most 16 digits.  67 looks
> rather odd...
>
Yes, it is only 16.  thank you so much.
>> +	writel_relaxed(val, info->base + info->offset);
> ... and you are promptly discarding the upper 32 bits, since writel_relaxed()
> takes u32:
> 	((void)__raw_writel((__force u32)cpu_to_le32(v),(c)))
> is going to truncate to 32bit, no matter what.  Quietly truncate, at that...
>
>> +const struct address_space_operations regfs_aops = {
>> +	.readpage   = simple_readpage,
>> +	.write_begin    = simple_write_begin,
>> +	.write_end  = simple_write_end,
>> +	.set_page_dirty = __set_page_dirty_buffers,
>> +};
> Again, huh?  What would use the page cache there, anyway?

actually it not need,  it will be remove

>> +static LIST_HEAD(regfs_head);
> Protected by...?
It is  better to add a protect for multpite mount, i will do more test 
for it.
>> +static const struct inode_operations regfs_dir_inode_operations;
>> +int regfs_debug;
>> +module_param(regfs_debug, int, S_IRUGO);
>> +MODULE_PARM_DESC(regfs_debug, "enable regfs debug mode");
>> +
>> +struct inode *regfs_get_inode(struct super_block *sb, const struct inode *dir, umode_t mode, dev_t dev)
>> +{
>> +	struct inode *inode = new_inode(sb);
>> +
>> +	if (inode) {
>> +		inode->i_ino = get_next_ino();
>> +		inode_init_owner(inode, dir, mode);
>> +		inode->i_mapping->a_ops = &regfs_aops;
>> +		//inode->i_mapping->backing_dev_info = &regfs_backing_dev_info;
>> +		mapping_set_gfp_mask(inode->i_mapping, GFP_HIGHUSER);
>> +		mapping_set_unevictable(inode->i_mapping);
>> +		inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
>> +		switch (mode & S_IFMT) {
>> +		default:
>> +			init_special_inode(inode, mode, dev);
>> +			break;
>> +		case S_IFREG:
>> +			inode->i_op = &regfs_file_inode_operations;
>> +			inode->i_fop = &regfs_file_operations;
>> +			break;
>> +		case S_IFDIR:
>> +			inode->i_op = &regfs_dir_inode_operations;
>> +			inode->i_fop = &simple_dir_operations;
>> +
>> +			/* directory inodes start off with i_nlink == 2 (for "." entry) */
>> +			inc_nlink(inode);
>> +			break;
>> +		case S_IFLNK:
>> +			inode->i_op = &page_symlink_inode_operations;
>> +			break;
>> +		}
>> +	}
>> +
>> +	return inode;
>> +}
> Seriously?  Where would symlinks, device nodes, FIFOs and sockets come from?
> And you are open-coding the regular file case in the new_dentry_create() anyway,
> so the only thing this is actually used for is the root directory.
yes,  i have limited the rm/create/link in regfs in inode_operations, so 
i need to remove all these and only leave the

inode->i_op = &regfs_file_inode_operations; ?

>> +static const struct inode_operations regfs_dir_inode_operations = {
>> +	.lookup		= simple_lookup,
>> +};
> ... and simple_dir_inode_operations is wrong, because...?
>
Here i limit the inode create/rm/link and so on, only support the read 
and write, it isn't useful for new inode,  because these inode have not 
register addr to bind its dentry,

i think these create/rm/simlink are invalidm,  only use simple_lookup.

>> +static struct dentry *new_dentry_create(struct super_block *sb, struct dentry *parent,
>> +		 const char *name, bool is_dir, struct res_data *res)
>> +{
>> +	struct dentry *dentry;
>> +	struct inode *inode;
>> +	struct regfs_inode_info *ei;
>> +	struct regfs_fs_info *fsi = sb->s_fs_info;
>> +
>> +	dentry = d_alloc_name(parent, name);
>> +	if (!dentry)
>> +		return NULL;
>> +
>> +	inode = new_inode(sb);
>> +	if (!inode)
>> +		goto out;
>> +
>> +	ei = REGFS_I(inode);
>> +	inode->i_ino = get_next_ino();;
>> +	inode->i_mtime = inode->i_atime = inode->i_ctime = current_time(inode);
>> +	inode->i_uid =  GLOBAL_ROOT_UID;
>> +	inode->i_gid =  GLOBAL_ROOT_GID;
>> +	if (is_dir) {
>> +		inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR;
>> +		inode->i_op = &regfs_dir_inode_operations;
>> +		inode->i_fop = &simple_dir_operations;
>> +		list_add(&ei->list, &fsi->list);
> where's the matching removal from the list?

Current regfs doesn't support  inode  create/del in regfs,  sorry i am 
lazy to free it,  all will be free when

umounting,  if inode can't del, it will never need to remove from list,  
but it is better to add the remove list.

>> +	} else {
>> +		inode->i_mode = S_IFREG | S_IRUGO | S_IWUSR;
>> +		inode->i_op = &regfs_file_inode_operations;
>> +		inode->i_fop = &regfs_file_operations;
>> +		inc_nlink(inode);
>> +	}
>> +	ei->base = (void *)res->base;
>> +	ei->offset = res->offset;
>> +	ei->type = res->type;
>> +
>> +	d_add(dentry, inode);
>> +
>> +	loc_debug("new dentry io base:%llx offset:%llx ei:%llx\n", (u64)ei->base, (u64)ei->offset, (u64)ei);
>> +	return dentry;
>> +out:
>> +	dput(dentry);
>> +	return NULL;
>> +}
>> +
>> +static void node_transfer_dentry(struct super_block *sb)
>> +{
>> +	struct regfs_fs_info *fsi = sb->s_fs_info;
>> +	void *blob = fsi->dtb_buf;
>> +	const char *pathp;
>> +	int node_offset, depth = -1;
>> +	struct dentry *parent = NULL;
>> +	u64 parent_base;
>> +
>> +	for (node_offset = fdt_next_node(blob, -1, &depth);
>> +		node_offset >= 0 && depth >= 0;
>> +		node_offset = fdt_next_node(blob, node_offset, &depth)) {
>> +
>> +		const struct fdt_property *prop;
>> +		struct res_data res;
>> +
>> +		pathp = fdt_get_name(blob, node_offset, NULL);
>> +		prop = (void *)fdt_getprop(blob, node_offset, "reg", NULL);
>> +
>> +		if (prop) {
>> +			unsigned long phys;
>> +
>> +			phys = fdt32_to_cpu(((const __be32 *)prop)[1]);
>> +			res.type = RES_TYPE_RANGE;
>> +			res.offset = fdt32_to_cpu(((const __be32 *)prop)[3]);
>> +			res.base = (u64)ioremap(phys, res.offset);
>> +
>> +			if (!res.base) {
>> +				parent = NULL;
>> +				parent_base = 0;
>> +				continue;
>> +			}
>> +
>> +			loc_debug("%s reg:%lx size:%lx map:%llx\n\n", pathp
>> +				 , (unsigned long) fdt32_to_cpu(((const __be32 *)prop)[1])
>> +				 , (unsigned long) fdt32_to_cpu(((const __be32 *)prop)[3])
>> +				 , (u64)res.base);
>> +
>> +			parent = new_dentry_create(sb, sb->s_root, (const char *)pathp, true, &res);
>> +			parent_base = res.base;
>> +
>> +		} else {
>> +			// parent dentry is create failed, igonre all child dentry
>> +			if (!parent)
>> +				continue;
>> +
>> +			prop = (void *)fdt_getprop(blob, node_offset, "offset", NULL);
>> +			if (prop) {
>> +
>> +				res.offset = fdt32_to_cpu(*(const __be32 *)prop);
>> +				res.base = parent_base;
>> +				res.type = RES_TYPE_ITEM;
>> +
>> +				new_dentry_create(sb, parent, (const char *) pathp, false, &res);
>> +				loc_debug("%s offset:%lx\n", pathp, (unsigned long)fdt32_to_cpu(*(const __be32 *)prop));
>> +			}
>> +		}
>> +	}
>> +}
>> +
>> +static int parse_options(char *options, struct super_block *sb)
>> +{
>> +	char *p;
>> +	int ret = -EINVAL;
>> +	struct regfs_fs_info *fsi;
>> +	size_t msize = INT_MAX;
>> +
>> +	fsi = sb->s_fs_info;
>> +
>> +	if (!options)
>> +		return -EINVAL;
>> +
>> +	while ((p = strsep(&options, ",")) != NULL) {
>> +		char *name, *name_val;
>> +
>> +		name = strsep(&p, "=");
>> +		if (name == NULL)
>> +			goto failed;
>> +
>> +		name_val = strsep(&p, "=");
>> +		if (name_val == NULL)
>> +			goto failed;
>> +
>> +		//get resource address
>> +		if (!strcmp(name, "dtb")) {
>> +			ret = kernel_read_file_from_path(name_val, &fsi->dtb_buf, &fsi->dtb_len, msize, READING_UNKNOWN);
> Why bother doing that in the kernel?

   Do you mean why using  kernel_read_file_from_path and bind these dtb 
into dentry?  i want to support different devices registers to mount

different path by using itself dtb file. i introduce mount param: 
dtb=test1.dtb,  load these dtb to module and explain them, later these 
registers

will bind into different  dentrys:

        mount -t regfs -o dtb=test.dtb none /mnt

     also you can mount others device dts to different dirtecotry.

        mount -t regfs -o dtb=test1.dtb none /mnt1

        mount -t regfs -o dtb=test2.dtb none /mnt2

I meet a hateful problem is devmem, for debugging some divices registers 
in user level,  I can't remember so many registers and to find their 
addr by read device spec,

if it can be list and more easy to write/read and don't let me to 
reconfig these map of dentry in next time,  i think dts file is suitably.

>> +struct dentry *regfs_mount(struct file_system_type *fs_type,
>> +	int flags, const char *dev_name, void *data)
>> +{
>> +	struct dentry *root_dentry;
>> +	struct super_block *sb;
>> +
>> +	root_dentry = mount_nodev(fs_type, flags, data, regfs_fill_super);
>> +
>> +	sb = root_dentry->d_sb;
>> +
>> +	if (sb->s_root) {
>> +		node_transfer_dentry(sb);
> Er... Why not do that in regfs_fill_super()?
Yes, it is betters to move it into regfs_fill_super.
> Al, not going any further for now...

the master problem is i need to support inode create, del and so on?  
event i provide these opereations, these new inodes haven't bind the dts 
register addr,

it just a page cache inode,  like ramfs inode, these inodes are invalid 
for regfs.