Re: [PATCH] vfs: preserve subsecond timestamps at filesystem time limits

[Jeff Layton]

On Sun, 2026-11-01 at 11:38 +0000, Hungyu Lin wrote:
> timestamp_truncate() clears tv_nsec whenever the clamped seconds equal
> either filesystem time limit. This also drops the fractional part of
> timestamps whose seconds were already within the supported range.
> 
> For example, setting a JFS mtime to 0.123456789 produces 0.000000000,
> although the filesystem can store the requested value. Utilities such
> as touch and cp -p consequently lose valid subsecond timestamps when
> preserving these endpoint values. Timestamps strictly between the
> filesystem limits are unaffected.
> 
> Only clear tv_nsec when clamping actually changes tv_sec. Continue to
> truncate in-range nanoseconds to the filesystem's timestamp granularity,
> including when the seconds equal either endpoint.
> 
> Ben Hutchings suggested comparing clamped and original seconds in the
> 2019 utimes review. That caller was switched to timestamp_truncate(),
> but the helper retained its endpoint equality check.
> 
> Link: https://lore.kernel.org/lkml/12c095e595836a7ff7f2c7b2a32cb5544dd29b55.camel@xxxxxxxxxxxxxxx/
> Fixes: 50e17c000c46 ("vfs: Add timestamp_truncate() api")
> Signed-off-by: Hungyu Lin <dennylin0707@xxxxxxxxx>
> ---
>  fs/inode.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/inode.c b/fs/inode.c
> index 1cb6293b237a..7f8e8feec4db 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -2847,9 +2847,10 @@ struct timespec64 timestamp_truncate(struct timespec64 t, struct inode *inode)
>  {
>  	struct super_block *sb = inode->i_sb;
>  	unsigned int gran = sb->s_time_gran;
> +	time64_t seconds = t.tv_sec;
>  
>  	t.tv_sec = clamp(t.tv_sec, sb->s_time_min, sb->s_time_max);
> -	if (unlikely(t.tv_sec == sb->s_time_max || t.tv_sec == sb->s_time_min))
> +	if (unlikely(t.tv_sec != seconds))
>  		t.tv_nsec = 0;
>  
>  	/* Avoid division in the common cases 1 ns and 1 s. */

Seems right. To wit: you only want to zero out tv_nsec in the case
where the clamp happens. If seconds was already at s_time_max/min,
don't zero out the nsec field.

Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>