Re: [PATCH v2 5/7] ext4: fix up test_dummy_encryption handling for new mount API

[Eric Biggers <ebiggers@kernel.org>]

On Fri, May 13, 2022 at 04:28:53PM +0530, Ritesh Harjani wrote:
> On 22/05/11 06:03PM, Eric Biggers wrote:
> > On Wed, May 11, 2022 at 11:24:33PM +0530, Ritesh Harjani wrote:
> > > On 22/05/09 04:40PM, Eric Biggers wrote:
> > > > A couple corrections I'll include in the next version:
> > >
> > > Need few clarifications. Could you please help explain what am I missing here?
> > >
> > > >
> > > > On Sat, Apr 30, 2022 at 10:08:55PM -0700, Eric Biggers wrote:
> > > > > +	if (fc->purpose == FS_CONTEXT_FOR_RECONFIGURE) {
> > > > > +		if (fscrypt_dummy_policies_equal(&sbi->s_dummy_enc_policy,
> > > > > +						 &ctx->dummy_enc_policy))
> > > > > +			return 0;
> > > > >  		ext4_msg(NULL, KERN_WARNING,
> > > > > -			 "Can't set test_dummy_encryption on remount");
> > > > > +			 "Can't set or change test_dummy_encryption on remount");
> > > > >  		return -EINVAL;
> > > > >  	}
> > > >
> > > > I think this needs to be 'fc->purpose == FS_CONTEXT_FOR_RECONFIGURE ||
> > > > fscrypt_is_dummy_policy_set(&sbi->s_dummy_enc_policy)', since ext4 can parse
> > > > mount options from both s_mount_opts and the regular mount options.
> > >
> > > Sorry, I am missing something here. Could you please help me understand why
> > > do we need the other OR case which you mentioned above i.e.
> > > "|| fscrypt_is_dummy_policy_set(&sbi->s_dummy_enc_policy)"
> > >
> > > So maybe to put it this way, when will it be the case where
> > > fscrypt_is_dummy_policy_set(&sbi->s_dummy_enc_policy) is true and it is not a
> > > FS_CONTEXT_FOR_RECONFIGURE case?
> >
> > The case where test_dummy_encryption is present in both the mount options stored
> > in the superblock and in the regular mount options.  See how __ext4_fill_super()
> > parses and applies each source of options separately.
> 
> Ok, thanks for clarifying. So this says that
> 1. in case of mount; if test_dummy_encryption is already set with some policy in
>    the disk superblock and if the user is trying to change the mount option in
>    options string, then that is not allowed.
> 2. Similarly if while remounting we try to change the mount option from the
>    previous mount option, then again this is not allowed.
> 

Yes.  I assume that the expected behavior of the on-disk mount options is the
same as if they were prepended to the user-specified mount options.  So we
simply aren't allowing conflicting test_dummy_encryption options in the mount
options, regardless of where the mount options came from.

> > > > > +static void ext4_apply_test_dummy_encryption(struct ext4_fs_context *ctx,
> > > > > +                                            struct super_block *sb)
> > > > > +{
> > > > > +	if (!fscrypt_is_dummy_policy_set(&ctx->dummy_enc_policy))
> > > > > +		return;
> > > >
> > > > To handle remounts correctly, this needs to be
> > > > '!fscrypt_is_dummy_policy_set(&ctx->dummy_enc_policy) ||
> > > > fscrypt_is_dummy_policy_set(&EXT4_SB(sb)->s_dummy_enc_policy)'.
> > >
> > > Why?
> > > Isn't it true that in remount we should update EXT4_SB(sb)->s_dummy_enc_policy
> > > only when ctx->dummy_enc_policy is set. If EXT4_SB(sb)->s_dummy_enc_policy is
> > > already set and ctx->dummy_enc_policy is not set, that means it's a remount case with no mount
> > > opts in which case ext4 should continue to have the same value of EXT4_SB(sb)->s_dummy_enc_policy?
> >
> > struct fscrypt_dummy_policy includes dynamically allocated memory, so
> > overwriting it without first freeing it would be a memory leak.
> 
> Ok yes. Since this is dynamic memory allocation. Hence
> I see that ext4_apply_test_dummy_encryption() can be called from
> parse_apply_sb_mount_options(), __ext4_fill_super() and __ext4_remount().
> 
> Case 1: when this mount option is set in superblock
> 1. So in parse_apply_sb_mount_options(), this mount option will get set the
>    first time if it is also set in superblock field.
> 
> 2. So if we also have a same mount option set in regular mount,
>    or during remount both will have sbi->s_dummy_enc_policy already set (from
>    step 1 above), so we should do nothing here.
> 
> Case 2: when this mount option is passed as regular mount
> 1. parse_apply_sb_mount_options() won't set this.
> 2. __ext4_fill_super() will set this mount option in sbi and hence __ext4_remount
>    should not set this again.
> 
> And as I see you are cleverly setting memset &ctx->dummy_enc_policy to 0
> in case where we applied the parsed mount option to sbi. So that the actual
> policy doesn't get free when you call __ext4_fc_free() after ext4_apply_options()
> in parse_apply_sb_mount_options(). And in other cases where this mount option was
> not applied to sbi mount opt, in that case we anyway want this policy to get
> free.
> 
> This somehow looks very confusing to me. But I guess with parse, check and apply
> mount APIs and with mount options in superblock, regular and remount path, this
> couldn't be avoided (although I am no expert in this area).
> 
> Thanks for explaining. I hope I got this right ;)

That's all correct.  I think you're overthinking it a bit.  The important thing
is that if the dummy policy is being set, we must move it into the ext4_sb_info.
Zeroing the old location is just part of transferring ownership of memory in C.
If a dummy policy was already set, we don't support changing it, and we've
checked that any "new" value is consistent with it, so we don't do anything.

- Eric