Re: [RESEND PATCH v9 1/3] libfs: Introduce case-insensitive string comparison helper

[Eugen Hristev]

On 2/8/24 20:38, Gabriel Krisman Bertazi wrote:
> Eugen Hristev <eugen.hristev@xxxxxxxxxxxxx> writes:
> 
>> From: Gabriel Krisman Bertazi <krisman@xxxxxxxxxxxxx>
>>
>> generic_ci_match can be used by case-insensitive filesystems to compare
>> strings under lookup with dirents in a case-insensitive way.  This
>> function is currently reimplemented by each filesystem supporting
>> casefolding, so this reduces code duplication in filesystem-specific
>> code.
>>
>> Reviewed-by: Eric Biggers <ebiggers@xxxxxxxxxx>
>> Signed-off-by: Gabriel Krisman Bertazi <krisman@xxxxxxxxxxxxx>
>> Signed-off-by: Eugen Hristev <eugen.hristev@xxxxxxxxxxxxx>
> 
> Hi Eugen,
> 
> Thanks for picking this up.  Please,  CC me in future versions.

Hello Gabriel,

Thanks for reviewing :)
> 
>> ---
>>  fs/libfs.c         | 68 ++++++++++++++++++++++++++++++++++++++++++++++
>>  include/linux/fs.h |  4 +++
>>  2 files changed, 72 insertions(+)
>>
>> diff --git a/fs/libfs.c b/fs/libfs.c
>> index bb18884ff20e..f80cb982ac89 100644
>> --- a/fs/libfs.c
>> +++ b/fs/libfs.c
>> @@ -1773,6 +1773,74 @@ static const struct dentry_operations generic_ci_dentry_ops = {
>>  	.d_hash = generic_ci_d_hash,
>>  	.d_compare = generic_ci_d_compare,
>>  };
>> +
>> +/**
>> + * generic_ci_match() - Match a name (case-insensitively) with a dirent.
>> + * @parent: Inode of the parent of the dirent under comparison
>> + * @name: name under lookup.
>> + * @folded_name: Optional pre-folded name under lookup
>> + * @de_name: Dirent name.
>> + * @de_name_len: dirent name length.
>> + *
>> + *
>> + * Test whether a case-insensitive directory entry matches the filename
>> + * being searched.  If @folded_name is provided, it is used instead of
>> + * recalculating the casefold of @name.
> 
> Can we add a note that this is a filesystem helper for comparison with
> directory entries, and VFS' ->d_compare should use generic_ci_d_compare.
> 
>> + *
>> + * Return: > 0 if the directory entry matches, 0 if it doesn't match, or
>> + * < 0 on error.
>> + */
>> +int generic_ci_match(const struct inode *parent,
>> +		     const struct qstr *name,
>> +		     const struct qstr *folded_name,
>> +		     const u8 *de_name, u32 de_name_len)
>> +{
>> +	const struct super_block *sb = parent->i_sb;
>> +	const struct unicode_map *um = sb->s_encoding;
>> +	struct fscrypt_str decrypted_name = FSTR_INIT(NULL, de_name_len);
>> +	struct qstr dirent = QSTR_INIT(de_name, de_name_len);
>> +	int res, match = false;
> 
> I know I originally wrote it this way, but match is an integer, so
> let's use integers instead of false/true.

With the changes below, 'match' is no longer needed
> 
>> +
>> +	if (IS_ENCRYPTED(parent)) {
>> +		const struct fscrypt_str encrypted_name =
>> +			FSTR_INIT((u8 *) de_name, de_name_len);
>> +
>> +		if (WARN_ON_ONCE(!fscrypt_has_encryption_key(parent)))
>> +			return -EINVAL;
>> +
>> +		decrypted_name.name = kmalloc(de_name_len, GFP_KERNEL);
>> +		if (!decrypted_name.name)
>> +			return -ENOMEM;
>> +		res = fscrypt_fname_disk_to_usr(parent, 0, 0, &encrypted_name,
>> +						&decrypted_name);
>> +		if (res < 0)
>> +			goto out;
>> +		dirent.name = decrypted_name.name;
>> +		dirent.len = decrypted_name.len;
>> +	}
>> +
>> +	if (folded_name->name)
>> +		res = utf8_strncasecmp_folded(um, folded_name, &dirent);
>> +	else
>> +		res = utf8_strncasecmp(um, name, &dirent);
> 
> Similar to
> 
>   https://git.kernel.org/pub/scm/linux/kernel/git/krisman/unicode.git/commit/?h=for-next&id=367122c529f35b4655acbe33c0cc4d6d3b32ba71
> 
> We should be checking for an exact-match first to avoid the utf8
> comparison cost unnecessarily.  The only problem is that we need to
> ensure we fail for an invalid utf-8 de_name in "strict mode".
> 
> Fortunately, if folded_name->name exists, we know the name-under-lookup
> was validated when initialized, so an exact-match of de_name must also
> be valid.  If folded_name is NULL, though, we might either have an
> invalid utf-8 dentry-under-lookup or the allocation itself failed, so we
> need to utf8_validate it.
> 
> Honestly, I don't care much about this !folded_name->name case, since
> utf8_strncasecmp will do the right thing and an invalid utf8 on
> case-insensitive directories should be an exception, not the norm.  but
> the code might get simpler if we do both:
> 
> (untested)

I implemented your suggestion, but any idea about testing ? I ran smoke on xfstests
and it appears to be fine, but maybe some specific test case might try the
different paths here ?

Let me know,
Eugen
> 
> if (folded_name->name) {
> 	if (dirent.len == folded_name->len &&
> 	    !memcmp(folded_name->name, dirent.name, dirent.len)) {
>             	res = 1;
> 		goto out;
>         }
> 	res = utf8_strncasecmp_folded(um, folded_name, &dirent);
> } else {
> 	if (dirent.len == name->len &&
> 	    !memcmp(name->name, dirent.name, dirent.len) &&
>             (!sb_has_strict_encoding(sb) || !utf8_validate(um, name))) {
>             	res = 1;
> 		goto out;
>         }
> 	res = utf8_strncasecmp(um, name, &dirent);
> }
> 
>> +
>> +	if (!res)
>> +		match = true;
>> +	else if (res < 0 && !sb_has_strict_encoding(sb)) {
>> +		/*
>> +		 * In non-strict mode, fallback to a byte comparison if
>> +		 * the names have invalid characters.
>> +		 */
>> +		res = 0;
>> +		match = ((name->len == dirent.len) &&
>> +			 !memcmp(name->name, dirent.name, dirent.len));
>> +	}
> 
> This goes away entirely.
> 
>> +
>> +out:
>> +	kfree(decrypted_name.name);
>> +	return (res >= 0) ? match : res;
> 
> and this becomes:
> 
> return res;
>