Re: [PATCH] lightnvm: pblk: adjust the position of the lock

[Matias BjÃrling]

On 11/27/2018 01:57 PM, Javier Gonzalez wrote:
> > On 27 Nov 2018, at 02.53, Hua Su <suhua.tanke@xxxxxxxxx> wrote:
> >
> > Add lock protection for list operations.
> > Signed-off-by: Hua Su <suhua.tanke@xxxxxxxxx>
> > ---
> > drivers/lightnvm/pblk-core.c | 17 ++++++++++-------
> > 1 file changed, 10 insertions(+), 7 deletions(-)
> >
> > diff --git a/drivers/lightnvm/pblk-core.c b/drivers/lightnvm/pblk-core.c
> > index 6944aac43b01..e490df217dac 100644
> > --- a/drivers/lightnvm/pblk-core.c
> > +++ b/drivers/lightnvm/pblk-core.c
> > @@ -1286,24 +1286,27 @@ int pblk_line_recov_alloc(struct pblk *pblk, struct pblk_line *line)
> > 	list_del(&line->list);
> >
> > 	ret = pblk_line_prepare(pblk, line);
> > -	if (ret) {
> > -		list_add(&line->list, &l_mg->free_list);
> > -		spin_unlock(&l_mg->free_lock);
> > -		return ret;
> > -	}
> > -	spin_unlock(&l_mg->free_lock);
> > +	if (ret)
> > +		goto out;
> >
> > 	ret = pblk_line_alloc_bitmaps(pblk, line);
> > 	if (ret)
> > -		return ret;
> > +		goto out;
> >
> > 	if (!pblk_line_init_bb(pblk, line, 0)) {
> > 		list_add(&line->list, &l_mg->free_list);
> > +		spin_unlock(&l_mg->free_lock);
> > 		return -EINTR;
> > 	}
> > +	spin_unlock(&l_mg->free_lock);
> >
> > 	pblk_rl_free_lines_dec(&pblk->rl, line, true);
> > 	return 0;
> > +
> > +out:
> > +	list_add(&line->list, &l_mg->free_list);
> > +	spin_unlock(&l_mg->free_lock);
> > +	return ret;
> > }
> >
> > void pblk_line_recov_close(struct pblk *pblk, struct pblk_line *line)
> > --
> > 2.19.1
>
> This path is only touched by the recovery path, which is single
> threaded, so there is no race condition as is. Also, if recovery fails,
> pblk will not create the instance at all. This said, it would be
> good to protect the list_add on the pblk_line_init_bb() error path in
> case this code is used for some other purpose in the future.

I like your explanation here. Another option is that we could add a 
comment to notify the developer that it safe in this context?

> However, the implementation you propose has some issues due to the
> extended critical zone, which covers pblk_line_alloc_bitmaps(). Here,
> allocations are intentionally non-atomic (i.e., GFP_KERNEL).
>
> You can instead do the following:
>
> diff --git i/drivers/lightnvm/pblk-core.c w/drivers/lightnvm/pblk-core.c
> index 615817bf97e3..02c4c1708c45 100644
> --- i/drivers/lightnvm/pblk-core.c
> +++ w/drivers/lightnvm/pblk-core.c
> @@ -1301,15 +1301,22 @@ int pblk_line_recov_alloc(struct pblk *pblk, struct pblk_line *line)
>
>          ret = pblk_line_alloc_bitmaps(pblk, line);
>          if (ret)
> -               return ret;
> +               goto fail;
>
>          if (!pblk_line_init_bb(pblk, line, 0)) {
> -               list_add(&line->list, &l_mg->free_list);
> -               return -EINTR;
> +               ret = -EINTR;
> +               goto fail;
>          }
>
>          pblk_rl_free_lines_dec(&pblk->rl, line, true);
>          return 0;
> +
> +fail:
> +       spin_lock(&l_mg->free_lock);
> +       list_add(&line->list, &l_mg->free_list);
> +       spin_unlock(&l_mg->free_lock);
> +
> +       return ret;
>   }
>
>   void pblk_line_recov_close(struct pblk *pblk, struct pblk_line *line)
>
> Javier