The cgwb cleanup routine will try to release the dying cgwb by switching
the attached inodes. It fetches the attached inodes from wb->b_attached
list, omitting the fact that inodes only with dirty timestamps reside in
wb->b_dirty_time list, which is the case when lazytime is enabled. This
causes enormous zombie memory cgroup when lazytime is enabled, as inodes
with dirty timestamps can not be switched to a live cgwb for a long time.
It is reasonable not to switch cgwb for inodes with dirty data, as
otherwise it may break the bandwidth restrictions. However since the
writeback of inode metadata is not accounted for, let's also switch
inodes with dirty timestamps to avoid zombie memory and block cgroups
when laztytime is enabled.
Fixs: c22d70a162d3 ("writeback, cgroup: release dying cgwbs by switching attached inodes")
Signed-off-by: Jingbo Xu <[email protected]>
---
v2: add comment explaining why switching for inodes with dirty
timestamps is needed
v1: https://lore.kernel.org/all/[email protected]/
---
fs/fs-writeback.c | 41 +++++++++++++++++++++++++++++------------
1 file changed, 29 insertions(+), 12 deletions(-)
diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index c1af01b2c42d..1767493dffda 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -613,6 +613,24 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
kfree(isw);
}
+static bool isw_prepare_wbs_switch(struct inode_switch_wbs_context *isw,
+ struct list_head *list, int *nr)
+{
+ struct inode *inode;
+
+ list_for_each_entry(inode, list, i_io_list) {
+ if (!inode_prepare_wbs_switch(inode, isw->new_wb))
+ continue;
+
+ isw->inodes[*nr] = inode;
+ (*nr)++;
+
+ if (*nr >= WB_MAX_INODES_PER_ISW - 1)
+ return true;
+ }
+ return false;
+}
+
/**
* cleanup_offline_cgwb - detach associated inodes
* @wb: target wb
@@ -625,7 +643,6 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
{
struct cgroup_subsys_state *memcg_css;
struct inode_switch_wbs_context *isw;
- struct inode *inode;
int nr;
bool restart = false;
@@ -647,17 +664,17 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
nr = 0;
spin_lock(&wb->list_lock);
- list_for_each_entry(inode, &wb->b_attached, i_io_list) {
- if (!inode_prepare_wbs_switch(inode, isw->new_wb))
- continue;
-
- isw->inodes[nr++] = inode;
-
- if (nr >= WB_MAX_INODES_PER_ISW - 1) {
- restart = true;
- break;
- }
- }
+ /*
+ * In addition to the inodes that have completed writeback, also switch
+ * cgwbs for those inodes only with dirty timestamps. Otherwise, those
+ * inodes won't be written back for a long time when lazytime is
+ * enabled, and thus pinning the dying cgwbs. It won't break the
+ * bandwidth restrictions, as writeback of inode metadata is not
+ * accounted for.
+ */
+ restart = isw_prepare_wbs_switch(isw, &wb->b_attached, &nr);
+ if (!restart)
+ restart = isw_prepare_wbs_switch(isw, &wb->b_dirty_time, &nr);
spin_unlock(&wb->list_lock);
/* no attached inodes? bail out */
--
2.19.1.6.gb485710b
On Fri 13-10-23 13:52:08, Jingbo Xu wrote:
> The cgwb cleanup routine will try to release the dying cgwb by switching
> the attached inodes. It fetches the attached inodes from wb->b_attached
> list, omitting the fact that inodes only with dirty timestamps reside in
> wb->b_dirty_time list, which is the case when lazytime is enabled. This
> causes enormous zombie memory cgroup when lazytime is enabled, as inodes
> with dirty timestamps can not be switched to a live cgwb for a long time.
>
> It is reasonable not to switch cgwb for inodes with dirty data, as
> otherwise it may break the bandwidth restrictions. However since the
> writeback of inode metadata is not accounted for, let's also switch
> inodes with dirty timestamps to avoid zombie memory and block cgroups
> when laztytime is enabled.
>
> Fixs: c22d70a162d3 ("writeback, cgroup: release dying cgwbs by switching attached inodes")
^^^ Fixes
> Signed-off-by: Jingbo Xu <[email protected]>
Otherwise looks good. Feel free to add:
Reviewed-by: Jan Kara <[email protected]>
Honza
> ---
> v2: add comment explaining why switching for inodes with dirty
> timestamps is needed
>
> v1: https://lore.kernel.org/all/[email protected]/
> ---
> fs/fs-writeback.c | 41 +++++++++++++++++++++++++++++------------
> 1 file changed, 29 insertions(+), 12 deletions(-)
>
> diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
> index c1af01b2c42d..1767493dffda 100644
> --- a/fs/fs-writeback.c
> +++ b/fs/fs-writeback.c
> @@ -613,6 +613,24 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
> kfree(isw);
> }
>
> +static bool isw_prepare_wbs_switch(struct inode_switch_wbs_context *isw,
> + struct list_head *list, int *nr)
> +{
> + struct inode *inode;
> +
> + list_for_each_entry(inode, list, i_io_list) {
> + if (!inode_prepare_wbs_switch(inode, isw->new_wb))
> + continue;
> +
> + isw->inodes[*nr] = inode;
> + (*nr)++;
> +
> + if (*nr >= WB_MAX_INODES_PER_ISW - 1)
> + return true;
> + }
> + return false;
> +}
> +
> /**
> * cleanup_offline_cgwb - detach associated inodes
> * @wb: target wb
> @@ -625,7 +643,6 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
> {
> struct cgroup_subsys_state *memcg_css;
> struct inode_switch_wbs_context *isw;
> - struct inode *inode;
> int nr;
> bool restart = false;
>
> @@ -647,17 +664,17 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
>
> nr = 0;
> spin_lock(&wb->list_lock);
> - list_for_each_entry(inode, &wb->b_attached, i_io_list) {
> - if (!inode_prepare_wbs_switch(inode, isw->new_wb))
> - continue;
> -
> - isw->inodes[nr++] = inode;
> -
> - if (nr >= WB_MAX_INODES_PER_ISW - 1) {
> - restart = true;
> - break;
> - }
> - }
> + /*
> + * In addition to the inodes that have completed writeback, also switch
> + * cgwbs for those inodes only with dirty timestamps. Otherwise, those
> + * inodes won't be written back for a long time when lazytime is
> + * enabled, and thus pinning the dying cgwbs. It won't break the
> + * bandwidth restrictions, as writeback of inode metadata is not
> + * accounted for.
> + */
> + restart = isw_prepare_wbs_switch(isw, &wb->b_attached, &nr);
> + if (!restart)
> + restart = isw_prepare_wbs_switch(isw, &wb->b_dirty_time, &nr);
> spin_unlock(&wb->list_lock);
>
> /* no attached inodes? bail out */
> --
> 2.19.1.6.gb485710b
>
--
Jan Kara <[email protected]>
SUSE Labs, CR
On Fri, Oct 13, 2023 at 01:52:08PM +0800, Jingbo Xu wrote:
> The cgwb cleanup routine will try to release the dying cgwb by switching
> the attached inodes. It fetches the attached inodes from wb->b_attached
> list, omitting the fact that inodes only with dirty timestamps reside in
> wb->b_dirty_time list, which is the case when lazytime is enabled. This
> causes enormous zombie memory cgroup when lazytime is enabled, as inodes
> with dirty timestamps can not be switched to a live cgwb for a long time.
>
> It is reasonable not to switch cgwb for inodes with dirty data, as
> otherwise it may break the bandwidth restrictions. However since the
> writeback of inode metadata is not accounted for, let's also switch
> inodes with dirty timestamps to avoid zombie memory and block cgroups
> when laztytime is enabled.
>
> Fixs: c22d70a162d3 ("writeback, cgroup: release dying cgwbs by switching attached inodes")
> Signed-off-by: Jingbo Xu <[email protected]>
Acked-by: Tejun Heo <[email protected]>
Thanks.
--
tejun