Re: [PATCH v16 3/9] libfs: Introduce case-insensitive string comparison helper

[Eugen Hristev <eugen.hristev@collabora.com>]

On 5/23/24 02:05, Gabriel Krisman Bertazi wrote:
> Eugen Hristev <eugen.hristev@collabora.com> writes:
> 
>> On 5/13/24 00:27, Gabriel Krisman Bertazi wrote:
>>> Eric Biggers <ebiggers@kernel.org> writes:
>>>
>>>> On Fri, Apr 05, 2024 at 03:13:26PM +0300, Eugen Hristev wrote:
>>>
>>>>> +		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;
>>>>
>>>> If fscrypt_fname_disk_to_usr() returns an error and !sb_has_strict_encoding(sb),
>>>> then this function returns 0 (indicating no match) instead of the error code
>>>> (indicating an error).  Is that the correct behavior?  I would think that
>>>> strict_encoding should only have an effect on the actual name
>>>> comparison.
>>>
>>> No. we *want* this return code to be propagated back to f2fs.  In ext4 it
>>> wouldn't matter since the error is not visible outside of ext4_match,
>>> but f2fs does the right thing and stops the lookup.
>>
>> In the previous version which I sent, you told me that the error should be
>> propagated only in strict_mode, and if !strict_mode, it should just return no match.
>> Originally I did not understand that this should be done only for utf8_strncasecmp
>> errors, and not for all the errors. I will change it here to fix that.
> 
> Yes, it depends on which error we are talking about. For ENOMEM and
> whatever error fscrypt_fname_disk_to_usr returns, we surely want to send
> that back, such that f2fs can handle it (i.e abort the lookup).  Unicode
> casefolding errors don't need to stop the lookup.
> 
> 
>>> Thinking about it, there is a second problem with this series.
>>> Currently, if we are on strict_mode, f2fs_match_ci_name does not
>>> propagate unicode errors back to f2fs. So, once a utf8 invalid sequence
>>> is found during lookup, it will be considered not-a-match but the lookup
>>> will continue.  This allows some lookups to succeed even in a corrupted
>>> directory.  With this patch, we will abort the lookup on the first
>>> error, breaking existing semantics.  Note that these are different from
>>> memory allocation failure and fscrypt_fname_disk_to_usr. For those, it
>>> makes sense to abort.
>>
>> So , in the case of f2fs , we must not propagate utf8 errors ? It should just
>> return no match even in strict mode ?
>> If this helper is common for both f2fs and ext4, we have to do the same for ext4 ?
>> Or we are no longer able to commonize the code altogether ?
> 
> We can have a common handler.  It doesn't matter for Ext4 because it
> ignores all errors. Perhaps ext4 can be improved too in a different
> patchset.
> 
>>> My suggestion would be to keep the current behavior.  Make
>>> generic_ci_match only propagate non-unicode related errors back to the
>>> filesystem.  This means that we need to move the error messages in patch
>>> 6 and 7 into this function, so they only trigger when utf8_strncasecmp*
>>> itself fails.
>>>
>>
>> So basically unicode errors stop here, and print the error message here in that case.
>> Am I understanding it correctly ?
> 
> Yes, that is it.  print the error message - only in strict mode - and
> return not-a-match.
> 
> Is there any problem with this approach that I'm missing?

As the printing is moved here, in the common code, we cannot use either of
f2fs_warn nor EXT4_ERROR_INODE . Any suggestion ? Would have to be something
meaningful for the user and ratelimited I guess.

Thanks for the explanations !


> 
>>>>> +	/*
>>>>> +	 * Attempt a case-sensitive match first. It is cheaper and
>>>>> +	 * should cover most lookups, including all the sane
>>>>> +	 * applications that expect a case-sensitive filesystem.
>>>>> +	 */
>>>>> +	if (folded_name->name) {
>>>>> +		if (dirent.len == folded_name->len &&
>>>>> +		    !memcmp(folded_name->name, dirent.name, dirent.len))
>>>>> +			goto out;
>>>>> +		res = utf8_strncasecmp_folded(um, folded_name, &dirent);
>>>>
>>>> Shouldn't the memcmp be done with the original user-specified name, not the
>>>> casefolded name?  I would think that the user-specified name is the one that's
>>>> more likely to match the on-disk name, because of case preservation.  In most
>>>> cases users will specify the same case on both file creation and later access.
>>>
>>> Yes.
>>>
>> so the utf8_strncasecmp_folded call here must use name->name instead of folded_name ?
> 
> No, utf8_strncasecmp_folded requires a casefolded name.  Eric's point is
> that the *memcmp* should always compare against name->name since it's more
> likely to match the name on disk than the folded version because the user
> is probably doing a case-exact lookup.
> 
> This also means the memcmp can be moved outside the "if (folded_name->name)",
> simplifying the patch!
>