| Message ID | 20190227171442.11853-4-igor.j.konopko@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | lightnvm: bugfixes and improvements | expand |
Looks good. Reviewed-by: Hans Holmberg <hans.holmberg@cnexlabs.com> On Wed, Feb 27, 2019 at 6:17 PM Igor Konopko <igor.j.konopko@intel.com> wrote: > > In current implementation of pblk_put_line_back behaviour > there are two cases which are not handled. > > First one is the race condition with __pblk_map_invalidate in > which function we check for line state, which might be closed, > but still not added to any list and thus explode in list_move_tail. > This is due to lack of locking both gc_lock and line lock > in pblk_put_line_back current implementation. > > The second issue is that when we are in that function, line > is not on any list and pblk_line_gc_list might hit the same > gc group and thus not return any move_list. Then our line > will stuck forever unassigned to any list. Simply resetting > gc_list to none will fix that. > > Signed-off-by: Igor Konopko <igor.j.konopko@intel.com> > --- > drivers/lightnvm/pblk-gc.c | 16 ++++++++++------ > 1 file changed, 10 insertions(+), 6 deletions(-) > > diff --git a/drivers/lightnvm/pblk-gc.c b/drivers/lightnvm/pblk-gc.c > index 31fc1339faa8..511ed0d5333c 100644 > --- a/drivers/lightnvm/pblk-gc.c > +++ b/drivers/lightnvm/pblk-gc.c > @@ -64,19 +64,23 @@ static void pblk_put_line_back(struct pblk *pblk, struct pblk_line *line) > struct pblk_line_mgmt *l_mg = &pblk->l_mg; > struct list_head *move_list; > > + spin_lock(&l_mg->gc_lock); > spin_lock(&line->lock); > WARN_ON(line->state != PBLK_LINESTATE_GC); > line->state = PBLK_LINESTATE_CLOSED; > trace_pblk_line_state(pblk_disk_name(pblk), line->id, > line->state); > + > + /* We need to reset gc_group in order to ensure that > + * pblk_line_gc_list will return proper move_list > + * since right now current line is not on any of the > + * gc lists. > + */ > + line->gc_group = PBLK_LINEGC_NONE; > move_list = pblk_line_gc_list(pblk, line); > spin_unlock(&line->lock); > - > - if (move_list) { > - spin_lock(&l_mg->gc_lock); > - list_add_tail(&line->list, move_list); > - spin_unlock(&l_mg->gc_lock); > - } > + list_add_tail(&line->list, move_list); > + spin_unlock(&l_mg->gc_lock); > } > > static void pblk_gc_line_ws(struct work_struct *work) > -- > 2.17.1 >
> On 27 Feb 2019, at 18.14, Igor Konopko <igor.j.konopko@intel.com> wrote: > > In current implementation of pblk_put_line_back behaviour > there are two cases which are not handled. > > First one is the race condition with __pblk_map_invalidate in > which function we check for line state, which might be closed, > but still not added to any list and thus explode in list_move_tail. > This is due to lack of locking both gc_lock and line lock > in pblk_put_line_back current implementation. > > The second issue is that when we are in that function, line > is not on any list and pblk_line_gc_list might hit the same > gc group and thus not return any move_list. Then our line > will stuck forever unassigned to any list. Simply resetting > gc_list to none will fix that. > > Signed-off-by: Igor Konopko <igor.j.konopko@intel.com> > --- > drivers/lightnvm/pblk-gc.c | 16 ++++++++++------ > 1 file changed, 10 insertions(+), 6 deletions(-) > > diff --git a/drivers/lightnvm/pblk-gc.c b/drivers/lightnvm/pblk-gc.c > index 31fc1339faa8..511ed0d5333c 100644 > --- a/drivers/lightnvm/pblk-gc.c > +++ b/drivers/lightnvm/pblk-gc.c > @@ -64,19 +64,23 @@ static void pblk_put_line_back(struct pblk *pblk, struct pblk_line *line) > struct pblk_line_mgmt *l_mg = &pblk->l_mg; > struct list_head *move_list; > > + spin_lock(&l_mg->gc_lock); > spin_lock(&line->lock); > WARN_ON(line->state != PBLK_LINESTATE_GC); > line->state = PBLK_LINESTATE_CLOSED; > trace_pblk_line_state(pblk_disk_name(pblk), line->id, > line->state); > + > + /* We need to reset gc_group in order to ensure that > + * pblk_line_gc_list will return proper move_list > + * since right now current line is not on any of the > + * gc lists. > + */ > + line->gc_group = PBLK_LINEGC_NONE; > move_list = pblk_line_gc_list(pblk, line); > spin_unlock(&line->lock); > - > - if (move_list) { > - spin_lock(&l_mg->gc_lock); > - list_add_tail(&line->list, move_list); > - spin_unlock(&l_mg->gc_lock); > - } > + list_add_tail(&line->list, move_list); > + spin_unlock(&l_mg->gc_lock); > } > > static void pblk_gc_line_ws(struct work_struct *work) > -- > 2.17.1 This comes back from the time where GC was single threaded - I left out this when reimplementing the locking. Good catch. Please, add a fixes tag for this to be back ported. Reviewed-by: Javier González <javier@javigon.com>
diff --git a/drivers/lightnvm/pblk-gc.c b/drivers/lightnvm/pblk-gc.c index 31fc1339faa8..511ed0d5333c 100644 --- a/drivers/lightnvm/pblk-gc.c +++ b/drivers/lightnvm/pblk-gc.c @@ -64,19 +64,23 @@ static void pblk_put_line_back(struct pblk *pblk, struct pblk_line *line) struct pblk_line_mgmt *l_mg = &pblk->l_mg; struct list_head *move_list; + spin_lock(&l_mg->gc_lock); spin_lock(&line->lock); WARN_ON(line->state != PBLK_LINESTATE_GC); line->state = PBLK_LINESTATE_CLOSED; trace_pblk_line_state(pblk_disk_name(pblk), line->id, line->state); + + /* We need to reset gc_group in order to ensure that + * pblk_line_gc_list will return proper move_list + * since right now current line is not on any of the + * gc lists. + */ + line->gc_group = PBLK_LINEGC_NONE; move_list = pblk_line_gc_list(pblk, line); spin_unlock(&line->lock); - - if (move_list) { - spin_lock(&l_mg->gc_lock); - list_add_tail(&line->list, move_list); - spin_unlock(&l_mg->gc_lock); - } + list_add_tail(&line->list, move_list); + spin_unlock(&l_mg->gc_lock); } static void pblk_gc_line_ws(struct work_struct *work)
In current implementation of pblk_put_line_back behaviour there are two cases which are not handled. First one is the race condition with __pblk_map_invalidate in which function we check for line state, which might be closed, but still not added to any list and thus explode in list_move_tail. This is due to lack of locking both gc_lock and line lock in pblk_put_line_back current implementation. The second issue is that when we are in that function, line is not on any list and pblk_line_gc_list might hit the same gc group and thus not return any move_list. Then our line will stuck forever unassigned to any list. Simply resetting gc_list to none will fix that. Signed-off-by: Igor Konopko <igor.j.konopko@intel.com> --- drivers/lightnvm/pblk-gc.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)