Re: [PATCH V5 0/9] Enable ext4 support for per-file/directory DAX operations

[Ira Weiny <ira.weiny@intel.com>]

On Fri, May 29, 2020 at 12:17:17AM -0400, Theodore Y. Ts'o wrote:
> On Thu, May 28, 2020 at 10:54:41PM -0400, Theodore Y. Ts'o wrote:
> > 
> > Thanks, applied to the ext4-dax branch.
> > 
> 
> I spoke too soon.  While I tried merging with the ext4.git dev branch,
> a merge conflict made me look closer and I realize I needed to make
> the following changes (see diff between your patch set and what is
> currently in ext4-dax).
> 
> Essentially, I needed to rework the branch to take into account commit
> e0198aff3ae3 ("ext4: reject mount options not supported when
> remounting in handle_mount_opt()").
> 
> The problem is that if you allow handle_mount_opt() to apply the
> changes to the dax settings, and then later on, ext4_remount() realize
> that we're remounting, and we need to reject the change, there's a
> race if we restore the mount options to the original configuration.
> Specifically, as Syzkaller pointed out, between when we change the dax
> settings and then reset them, it's possible for some file to be opened
> with "wrong" dax setting, and then when they are reset, *boom*.
> 
> The correct way to deal with this is to reject the mount option change
> much earlier, in handle_mount_opt(), *before* we mess with the dax
> settings.
> 
> Please take a look at the ext4-dax for the actual changes which I
> made.
> 
> Cheers,
> 
> 					- Ted
> 
> 
> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
> index 3658e3016999..9a37d70394b2 100644
> --- a/fs/ext4/super.c
> +++ b/fs/ext4/super.c
> @@ -1733,7 +1733,7 @@ static int clear_qf_name(struct super_block *sb, int qtype)
>  #define MOPT_NO_EXT3	0x0200
>  #define MOPT_EXT4_ONLY	(MOPT_NO_EXT2 | MOPT_NO_EXT3)
>  #define MOPT_STRING	0x0400
> -#define MOPT_SKIP	0x0800

I think we still need MOPT_SKIP...

This was put in to skip these options when printing to deal with printing only
dax=inode when it was specified by the user.

Ah but I see now.  By taking MOPT_SET away you have created the same behavior?

This is  orthogonal to the remount issue right?

> +#define MOPT_NO_REMOUNT	0x0800
>  
>  static const struct mount_opts {
>  	int	token;
> @@ -1783,18 +1783,15 @@ static const struct mount_opts {
>  	{Opt_min_batch_time, 0, MOPT_GTE0},
>  	{Opt_inode_readahead_blks, 0, MOPT_GTE0},
>  	{Opt_init_itable, 0, MOPT_GTE0},
> -	{Opt_dax, EXT4_MOUNT_DAX_ALWAYS, MOPT_SET | MOPT_SKIP},
> -	{Opt_dax_always, EXT4_MOUNT_DAX_ALWAYS,
> -		MOPT_EXT4_ONLY | MOPT_SET | MOPT_SKIP},
> -	{Opt_dax_inode, EXT4_MOUNT2_DAX_INODE,
> -		MOPT_EXT4_ONLY | MOPT_SET | MOPT_SKIP},
> -	{Opt_dax_never, EXT4_MOUNT2_DAX_NEVER,
> -		MOPT_EXT4_ONLY | MOPT_SET | MOPT_SKIP},
> +	{Opt_dax, 0, MOPT_NO_REMOUNT},
> +	{Opt_dax_always, 0, MOPT_NO_REMOUNT},
> +	{Opt_dax_inode, 0, MOPT_NO_REMOUNT},
> +	{Opt_dax_never, 0, MOPT_NO_REMOUNT},

Even if MOPT_SET is redundant.  Why don't we need still need MOPT_EXT4_ONLY?

And why don't we need to associate the defines; EXT4_MOUNT_DAX_ALWAYS etc?

>  	{Opt_stripe, 0, MOPT_GTE0},
>  	{Opt_resuid, 0, MOPT_GTE0},
>  	{Opt_resgid, 0, MOPT_GTE0},
> -	{Opt_journal_dev, 0, MOPT_NO_EXT2 | MOPT_GTE0},
> -	{Opt_journal_path, 0, MOPT_NO_EXT2 | MOPT_STRING},
> +	{Opt_journal_dev, 0, MOPT_NO_EXT2 | MOPT_GTE0 | MOPT_NO_REMOUNT},
> +	{Opt_journal_path, 0, MOPT_NO_EXT2 | MOPT_STRING | MOPT_NO_REMOUNT},
>  	{Opt_journal_ioprio, 0, MOPT_NO_EXT2 | MOPT_GTE0},
>  	{Opt_data_journal, EXT4_MOUNT_JOURNAL_DATA, MOPT_NO_EXT2 | MOPT_DATAJ},
>  	{Opt_data_ordered, EXT4_MOUNT_ORDERED_DATA, MOPT_NO_EXT2 | MOPT_DATAJ},
> @@ -1831,7 +1828,7 @@ static const struct mount_opts {
>  	{Opt_jqfmt_vfsv1, QFMT_VFS_V1, MOPT_QFMT},
>  	{Opt_max_dir_size_kb, 0, MOPT_GTE0},
>  	{Opt_test_dummy_encryption, 0, MOPT_GTE0},
> -	{Opt_nombcache, EXT4_MOUNT_NO_MBCACHE, MOPT_SET},
> +	{Opt_nombcache, EXT4_MOUNT_NO_MBCACHE, MOPT_SET | MOPT_NO_REMOUNT},
>  	{Opt_err, 0, 0}
>  };
>  
> @@ -1929,6 +1926,12 @@ static int handle_mount_opt(struct super_block *sb, char *opt, int token,
>  			 "Mount option \"%s\" incompatible with ext3", opt);
>  		return -1;
>  	}
> +	if ((m->flags & MOPT_NO_REMOUNT) && is_remount) {
> +		ext4_msg(sb, KERN_ERR,
> +			 "Mount option \"%s\" not supported when remounting",
> +			 opt);
> +		return -1;
> +	}

I think this is cleaner!

Thanks, I did test this but not while trying to manipulate files as the same time
as a remount.  So a race would not have been caught.

Thanks!
Ira

>  
>  	if (args->from && !(m->flags & MOPT_STRING) && match_int(args, &arg))
>  		return -1;
> @@ -2008,11 +2011,6 @@ static int handle_mount_opt(struct super_block *sb, char *opt, int token,
>  		}
>  		sbi->s_resgid = gid;
>  	} else if (token == Opt_journal_dev) {
> -		if (is_remount) {
> -			ext4_msg(sb, KERN_ERR,
> -				 "Cannot specify journal on remount");
> -			return -1;
> -		}
>  		*journal_devnum = arg;
>  	} else if (token == Opt_journal_path) {
>  		char *journal_path;
> @@ -2020,11 +2018,6 @@ static int handle_mount_opt(struct super_block *sb, char *opt, int token,
>  		struct path path;
>  		int error;
>  
> -		if (is_remount) {
> -			ext4_msg(sb, KERN_ERR,
> -				 "Cannot specify journal on remount");
> -			return -1;
> -		}
>  		journal_path = match_strdup(&args[0]);
>  		if (!journal_path) {
>  			ext4_msg(sb, KERN_ERR, "error: could not dup "
> @@ -2287,7 +2280,7 @@ static int _ext4_show_options(struct seq_file *seq, struct super_block *sb,
>  	for (m = ext4_mount_opts; m->token != Opt_err; m++) {
>  		int want_set = m->flags & MOPT_SET;
>  		if (((m->flags & (MOPT_SET|MOPT_CLEAR)) == 0) ||
> -		    (m->flags & MOPT_CLEAR_ERR) || m->flags & MOPT_SKIP)
> +		    (m->flags & MOPT_CLEAR_ERR))
>  			continue;
>  		if (!nodefs && !(m->mount_opt & (sbi->s_mount_opt ^ def_mount_opt)))
>  			continue; /* skip if same as the default */
> @@ -5474,24 +5467,6 @@ static int ext4_remount(struct super_block *sb, int *flags, char *data)
>  		}
>  	}
>  
> -	if ((sbi->s_mount_opt ^ old_opts.s_mount_opt) & EXT4_MOUNT_NO_MBCACHE) {
> -		ext4_msg(sb, KERN_ERR, "can't enable nombcache during remount");
> -		err = -EINVAL;
> -		goto restore_opts;
> -	}
> -
> -	if ((sbi->s_mount_opt ^ old_opts.s_mount_opt) & EXT4_MOUNT_DAX_ALWAYS ||
> -	    (sbi->s_mount_opt2 ^ old_opts.s_mount_opt2) & EXT4_MOUNT2_DAX_NEVER ||
> -	    (sbi->s_mount_opt2 ^ old_opts.s_mount_opt2) & EXT4_MOUNT2_DAX_INODE) {
> -		ext4_msg(sb, KERN_WARNING, "warning: refusing change of "
> -			"dax mount option with busy inodes while remounting");
> -		sbi->s_mount_opt &= ~EXT4_MOUNT_DAX_ALWAYS;
> -		sbi->s_mount_opt |= old_opts.s_mount_opt & EXT4_MOUNT_DAX_ALWAYS;
> -		sbi->s_mount_opt2 &= ~(EXT4_MOUNT2_DAX_NEVER | EXT4_MOUNT2_DAX_INODE);
> -		sbi->s_mount_opt2 |= old_opts.s_mount_opt2 &
> -				     (EXT4_MOUNT2_DAX_NEVER | EXT4_MOUNT2_DAX_INODE);
> -	}
> -
>  	if (sbi->s_mount_flags & EXT4_MF_FS_ABORTED)
>  		ext4_abort(sb, EXT4_ERR_ESHUTDOWN, "Abort forced by user");
>