On Fri, Oct 13, 2023 at 02:20:23PM -0700, Bart Van Assche wrote:
> On 10/13/23 02:33, Niklas Cassel wrote:
> > In commit c75e707fe1aa ("block: remove the per-bio/request write hint")
> > this line from fs/direct-io.c was removed:
> > - bio->bi_write_hint = dio->iocb->ki_hint;
> >
> > I'm not sure why this series does not readd a similar line to set the
> > lifetime (using bio_set_data_lifetime()) also for fs/direct-io.c.
>
> It depends on how we want the user to specify the data lifetime for
> direct I/O. This assignment is not modified by this patch series and
> copies the data lifetime information from the ioprio bitfield from user
> space into the bio:
>
> bio->bi_ioprio = dio->iocb->ki_ioprio;
Before per-bio/request write hints were removed, things looked like this:
io_uring.c:
req->rw.kiocb.ki_hint = ki_hint_validate(file_write_hint(req->file));
fs/fcntl.c:
static inline enum rw_hint file_write_hint(struct file *file)
{
if (file->f_write_hint != WRITE_LIFE_NOT_SET)
return file->f_write_hint;
return file_inode(file)->i_write_hint;
}
direct-io.c:
bio->bi_write_hint = dio->iocb->ki_hint;
buffered-io.c:
bio->bi_write_hint = inode->i_write_hint;
After this series, things instead look like this:
direct-io.c:
bio->bi_ioprio = dio->iocb->ki_ioprio;
buffered-io.c:
bio_set_data_lifetime(bio, inode->i_write_hint);
So when you say:
"It depends on how we want the user to specify the data lifetime for
direct I/O.", do you mean that buffered I/O should use fcntl() to specify
data lifetime, but direct I/O should used Linux IO priority API to specify
the same?
Because, to me that seems to be how the series is currently working.
(I am sorry if I am missing something.)
Kind regards,
Niklas
