On Sun, Jul 26, 2020 at 10:57:26AM -0400, Yafang Shao wrote:
> Bellow comment is quoted from Dave,
FYI, you mean "Below", not "Bellow". Dave doesn't often bellow.
> As a result, we should reintroduce PF_FSTRANS. Because PF_FSTRANS is only
> set by current, we can move it out of task->flags to avoid being out of PF_
> flags. So a new flag in_fstrans is introduced.
I don't think we need a new flag for this. I think you can just set
current->journal_info to a non-NULL value.
> +++ b/fs/xfs/xfs_linux.h
> @@ -111,6 +111,20 @@ typedef __u32 xfs_nlink_t;
> #define current_restore_flags_nested(sp, f) \
> (current->flags = ((current->flags & ~(f)) | (*(sp) & (f))))
>
> +static inline unsigned int xfs_trans_context_start(void)
> +{
> + unsigned int flags = current->in_fstrans;
> +
> + current->in_fstrans = 1;
> +
> + return flags;
> +}
> +
> +static inline void xfs_trans_context_end(unsigned int flags)
> +{
> + current->in_fstrans = flags ? 1 : 0;
> +}
Does XFS support nested transactions? If we're just using
current->journal_info, we can pretend its an unsigned long and use it
as a counter rather than handle the nesting the same way as the GFP flags.
