Re: [PATCH 2/3] ext4: implement IOCB_DONTCACHE handling in write operations

["Theodore Ts'o" <tytso@mit.edu>]

On Mon, Apr 21, 2025 at 10:50:30AM +0000, 陈涛涛 Taotao Chen wrote:
> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> index 94c7d2d828a6..787dd152a47e 100644
> --- a/fs/ext4/inode.c
> +++ b/fs/ext4/inode.c
> @@ -1147,16 +1147,22 @@ static int ext4_write_begin(struct file *file, struct address_space *mapping,
>  {
>  	struct inode *inode = mapping->host;
>  	int ret, needed_blocks;
> +	int iocb_flag;
>  	handle_t *handle;
>  	int retries = 0;
>  	struct folio *folio;
>  	pgoff_t index;
> +	fgf_t fgp = FGP_WRITEBEGIN;
>  	unsigned from, to;
>  
>  	ret = ext4_emergency_state(inode->i_sb);
>  	if (unlikely(ret))
>  		return ret;
>  
> +	iocb_flag = (int)(uintptr_t)(*fsdata);
> +	if (iocb_flag & IOCB_DONTCACHE)
> +		fgp |= FGP_DONTCACHE;
> +

See my comment against the first patch in this series.  It *should* be
possible to solve the problem just for ext4 by adding this line here:

	*fsdata = (void *)0;

The problem is that it's super-fragile, since how *fsdata gets used
changes at different points in time, so it makes code review and
maintenance more difficult.  (As evidenced by the fact that you missed
this; this is not a criticism on your programming ability, but rather
for the design choise of overloading the use of *fsdata.  This is a
trap that someone else might fall into when doing future code
changes.)

And of course, the question is whether PATCH 1/3 could potentially
break other file systems.  We would need audit all of the other
*_write_begin() functions, and then document this for the sake of
future file system developers that might want to change their
write_begin() function.

This is why my preference would be to add an extra flags paramter to
write_begin(), but that is going to be a lot more work.

Cheers,

						- Ted