mm: start background writeback based on per-wb threshold for strictlimit BDIs
Subsystems (3)
cover mm: start background writeback based on per-wb threshold for strictlimit BDIs
This patch has a dependency on the writeback regression fix in [1]. This is submitted as a separate patchset from [1] because [1] is a fix targeting the 7.0 rc cycle while this patch is an optimization that targets the next merge window.
Thanks,
Joanne
[1] https://lore.kernel.org/linux-fsdevel/[email protected]/
Joanne Koong (1):
mm: start background writeback based on per-wb threshold for
strictlimit BDIs
mm/page-writeback.c | 16 ++++++---------- 1 file changed, 6 insertions(+), 10 deletions(-)
Signature · 1 line
2.52.0
PATCH 1/1 mm: start background writeback based on per-wb threshold for strictlimit BDIs
The proactive nr_dirty > gdtc->bg_thresh check in balance_dirty_pages() only checks the global dirty threshold to start background writeback while the writer is still free-running, but for strictlimit BDIs (eg fuse), the per-wb dirty count can exceed the per-wb background threshold while the global threshold is not yet exceeded, so background writeback for this case never gets proactively started.
Add a per-wb threshold check for strictlimit BDIs so that background writeback is started when wb_dirty exceeds wb_bg_thresh, which drains dirty pages before the writer hits the throttle wall, matching the proactive behavior that the global check provides for non-strictlimit BDIs.
fio runs on fuse show about a 3-4% improvement in perf for buffered writes:
fio --name=writeback_test --ioengine=psync --rw=write --bs=128k \
--size=2G --numjobs=4 --ramp_time=10 --runtime=20 \
--time_based --group_reporting=1 --direct=0--- mm/page-writeback.c | 16 ++++++---------- 1 file changed, 6 insertions(+), 10 deletions(-)
mm/page-writeback.c +6 -10
@@ -1835,7 +1835,9 @@ static int balance_dirty_pages(struct bdi_writeback *wb,balance_domain_limits (mdtc, strictlimit); } -if (nr_dirty > gdtc->bg_thresh && !writeback_in_progress (wb)) +if (!writeback_in_progress (wb) && + (nr_dirty > gdtc->bg_thresh || + (strictlimit && gdtc->wb_dirty > gdtc->wb_bg_thresh)))wb_start_background_writeback (wb);/*
@@ -1862,15 +1864,9 @@ static int balance_dirty_pages(struct bdi_writeback *wb,* Unconditionally start background writeback if it's not * already in progress. We need to do this because the global * dirty threshold check above (nr_dirty > gdtc->bg_thresh) -* doesn't account for these cases: -* -* a) strictlimit BDIs: throttling is calculated using per-wb -* thresholds. The per-wb threshold can be exceeded even when -* nr_dirty < gdtc->bg_thresh -* -* b) memcg-based throttling: memcg uses its own dirty count and -* thresholds and can trigger throttling even when global -* nr_dirty < gdtc->bg_thresh +* doesn't account for the memcg-based throttling case. memcg +* uses its own dirty count and thresholds and can trigger +* throttling even when global nr_dirty < gdtc->bg_thresh * * Writeback needs to be started else the writer stalls in the * throttle loop waiting for dirty pages to be written back
Signature · 1 line
2.52.0
Andrew Morton
maintainerQuoted 13 lines · On Thu, 26 Mar 2026 16:46:29 -0700 Joanne Koong <[email protected]> wrote:
The proactive
nr_dirty>gdtc->bg_threshcheck inbalance_dirty_pages()only checks the global dirty threshold to start background writeback while the writer is still free-running, but for strictlimit BDIs (eg fuse), the per-wb dirty count can exceed the per-wb background threshold while the global threshold is not yet exceeded, so background writeback for this case never gets proactively started.Add a per-wb threshold check for strictlimit BDIs so that background writeback is started when
wb_dirtyexceedswb_bg_thresh, which drains dirty pages before the writer hits the throttle wall, matching the proactive behavior that the global check provides for non-strictlimit BDIs.
Quoted 6 lines
fio runs on fuse show about a 3-4% improvement in perf for buffered writes:
fio --name=writeback_test --ioengine=psync --rw=write --bs=128k \ --size=2G --numjobs=4 --ramp_time=10 --runtime=20 \ --time_based --group_reporting=1 --direct=0
Thanks.
Your v1 patch's changelog had useful information showing the performance impact on fuse. I've pasted one sentence from that changelog into my copy of this changelog, but please don't be sparing in describing the importance of a fix.
Also, we lost the Fixes tag. I've added
to my copy of this v2 patch. Please confirm that this is appropriate.
Quoted 11 lines
mm/page-writeback.c +0
-5@@ -1862,15 +1864,9 @@ static int balance_dirty_pages(struct bdi_writeback *wb,* Unconditionally start background writeback if it's not * already in progress. We need to do this because the global * dirty threshold check above (nr_dirty > gdtc->bg_thresh) -* doesn't account for these cases: -* -* a) strictlimit BDIs: throttling is calculated using per-wb -* thresholds. The per-wb threshold can be exceeded even when -* nr_dirty < gdtc->bg_thresh
mm/page-writeback.c static int balance_dirty_pages(struct bdi_writeback *wb, · line 1872
Quoted 10 lines
mm/page-writeback.c +3
-4-* -* b) memcg-based throttling: memcg uses its own dirty count and -* thresholds and can trigger throttling even when global -* nr_dirty < gdtc->bg_thresh +* doesn't account for the memcg-based throttling case. memcg +* uses its own dirty count and thresholds and can trigger +* throttling even when global nr_dirty < gdtc->bg_thresh * * Writeback needs to be started else the writer stalls in the * throttle loop waiting for dirty pages to be written back
My copy of balance_dirty_pages() doesn't have this comment block. I assume you're patching a different kernel version. I dropped this hunk. Please check current Linus mainline and see if there's somewhere else where you'd like to add this information.
Andrew Morton
maintainerQuoted 11 lines · On Thu, 26 Mar 2026 17:42:37 -0700 Andrew Morton <[email protected]>…
Quoted 6 lines
fio runs on fuse show about a 3-4% improvement in perf for buffered writes:
fio --name=writeback_test --ioengine=psync --rw=write --bs=128k \ --size=2G --numjobs=4 --ramp_time=10 --runtime=20 \ --time_based --group_reporting=1 --direct=0Thanks.
[bunch of nonsense]
argh, sorry, I didn't read your [0/N]. Ignore.
btw, a [0/N] isn't really needed for a singleton patch - best to just put everything into the patch's changelog.
Transient info such as "this depends on my other patch" can be placed below the singleton patch's "---" separator.
Quoted 8 lines · Andrew Morton
On Thu, 26 Mar 2026 17:42:37 -0700 Andrew Morton <[email protected]> wrote:
btw, a [0/N] isn't really needed for a singleton patch - best to just put everything into the patch's changelog.
Transient info such as "this depends on my other patch" can be placed below the singleton patch's "---" separator.
Ah I see, I will do that instead next time, sorry for the confusion.
Thanks,
Joanne
Quoted 13 lines · Joanne Koong
The proactive
nr_dirty>gdtc->bg_threshcheck inbalance_dirty_pages()only checks the global dirty threshold to start background writeback while the writer is still free-running, but for strictlimit BDIs (eg fuse), the per-wb dirty count can exceed the per-wb background threshold while the global threshold is not yet exceeded, so background writeback for this case never gets proactively started.Add a per-wb threshold check for strictlimit BDIs so that background writeback is started when
wb_dirtyexceedswb_bg_thresh, which drains dirty pages before the writer hits the throttle wall, matching the proactive behavior that the global check provides for non-strictlimit BDIs.
Quoted 7 lines
fio runs on fuse show about a 3-4% improvement in perf for buffered writes:
fio --name=writeback_test --ioengine=psync --rw=write --bs=128k \ --size=2G --numjobs=4 --ramp_time=10 --runtime=20 \ --time_based --group_reporting=1 --direct=0Signed-off-by Joanne Koong <[email protected]>
Looks good. Thanks for improving these! Feel free to add:
Honza
Quoted 41 lines
--- mm/page-writeback.c | 16 ++++++---------- 1 file changed, 6 insertions(+), 10 deletions(-)mm/page-writeback.c +6
-10@@ -1835,7 +1835,9 @@ static int balance_dirty_pages(struct bdi_writeback *wb,balance_domain_limits (mdtc, strictlimit); } -if (nr_dirty > gdtc->bg_thresh && !writeback_in_progress (wb)) +if (!writeback_in_progress (wb) && + (nr_dirty > gdtc->bg_thresh || + (strictlimit && gdtc->wb_dirty > gdtc->wb_bg_thresh)))wb_start_background_writeback (wb);/* @@ -1862,15 +1864,9 @@ static int balance_dirty_pages(struct bdi_writeback *wb,* Unconditionally start background writeback if it's not * already in progress. We need to do this because the global * dirty threshold check above (nr_dirty > gdtc->bg_thresh) -* doesn't account for these cases: -* -* a) strictlimit BDIs: throttling is calculated using per-wb -* thresholds. The per-wb threshold can be exceeded even when -* nr_dirty < gdtc->bg_thresh -* -* b) memcg-based throttling: memcg uses its own dirty count and -* thresholds and can trigger throttling even when global -* nr_dirty < gdtc->bg_thresh +* doesn't account for the memcg-based throttling case. memcg +* uses its own dirty count and thresholds and can trigger +* throttling even when global nr_dirty < gdtc->bg_thresh * * Writeback needs to be started else the writer stalls in the * throttle loop waiting for dirty pages to be written back Signature · 1 line
2.52.0
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR