`memory.stat:writeback` already gives you the exact real-time count of
pages still waiting to finish write-back for that cgroup, and it is
updated atomically in the hot path (`set_page_writeback()` /
`end_page_writeback()`). Reading it is just an `atomic_long_read()` (or
per-cpu equivalent), so the extra CPU cost of exposing it is essentially
zero. I was thinking that extra additional info would help us
Thanks,
Giorgi
On 8/21/2025 6:37 AM, Julian Sun wrote:
Hi, thanks for your review.
On Wed, Aug 20, 2025 at 8:17 PM Giorgi Tchankvetadze
<giorgitchankvetadze1997@xxxxxxxxx> wrote:
Could we add wb_pending_pages to memory.events?
Very cheap and useful.
A single atomic counter is already kept internally; exposing it is one
line in memcontrol.c plus one line in the ABI doc.
Not sure what do you mean by wb_pending_pages? Another counter besides
existing MEMCG_LOW MEMCG_HIGH MEMCG_MAX, etc.? And AFAIK there's no
pending pages in this patch set. Could you give more details?
Thanks,
On 8/20/2025 3:19 PM, Julian Sun wrote:
This patch series aims to eliminate task hangs in mem_cgroup_css_free()
caused by calling wb_wait_for_completion().
This is because there may be a large number of writeback tasks in the
foreign memcg, involving millions of pages, and the situation is
exacerbated by WBT rate limiting—potentially leading to task hangs
lasting several hours.
Patch 1 is preparatory work and involves no functional changes.
Patch 2 implements the automatic release of wb_completion.
Patch 3 removes wb_wait_for_completion() from mem_cgroup_css_free().
Julian Sun (3):
writeback: Rename wb_writeback_work->auto_free to free_work.
writeback: Add wb_writeback_work->free_done
memcg: Don't wait writeback completion when release memcg.
fs/fs-writeback.c | 22 ++++++++++++++--------
include/linux/backing-dev-defs.h | 6 ++++++
include/linux/memcontrol.h | 2 +-
mm/memcontrol.c | 29 ++++++++++++++++++++---------
4 files changed, 41 insertions(+), 18 deletions(-)
