Re: [PATCH 1/3] mm/filemap: initialize fsdata with iocb->ki_flags

["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/mm/filemap.c b/mm/filemap.c
> index 7b90cbeb4a1a..9174b6310f0b 100644
> --- a/mm/filemap.c
> +++ b/mm/filemap.c
> @@ -4087,7 +4087,11 @@ ssize_t generic_perform_write(struct kiocb *iocb, struct iov_iter *i)
>  		size_t offset;		/* Offset into folio */
>  		size_t bytes;		/* Bytes to write to folio */
>  		size_t copied;		/* Bytes copied from user */
> -		void *fsdata = NULL;
> +		/*
> +		 * Initialize fsdata with iocb flags pointer so that filesystems
> +		 * can check ki_flags (like IOCB_DONTCACHE) in write operations.
> +		 */
> +		void *fsdata = &iocb->ki_flags;

Unfortunately, this is't safe.  There may very well be code
paths which depend on fsdata being initialized to NULL before
calling write_begin().

In fact in the patch 2/3 in this series, ext4_write_end() will get
confused in the non-delayed allocation case, since ext4_write_begin()
doesn't force fsdata to be not be &iocb->ki_flags, and this will cause
ext4_write_end() to potentially get confused and do the wrong thing.

I understand that it would be a lot more inconvenient change the
function signature of write_begin() to pass through iocb->ki_fags via
a new parameter.  But I think that probably is the best way to go.

Cheers,

					- Ted