[PATCH v6 0/5] mm: batch TLB flushing for dirty folios in vmscan
From: Barry Song (Xiaomi)
Date: Sun Nov 15 2026 - 18:02:44 EST
On Sat, Oct 31, 2026 at 6:35 PM Zhang Peng <zippermonkey@xxxxxxxxxx> wrote:
>
> This series introduces batch TLB flushing optimization for dirty folios
> during memory reclaim, aiming to reduce IPI overhead on multi-core systems.
>
> Background
> ----------
> Reclaim calls try_to_unmap_flush_dirty() before writing out each dirty
> folio. When writable TLB entries are pending, this triggers a flush.
> On systems that use IPIs for TLB shootdowns, frequent flushes can add
> significant overhead.
>
> Approach
> --------
> Collect dirty folios into a folio_batch, releasing their locks while
> collecting the batch so swap faults can progress. Before writeout, try
> locking each folio again and recheck writeback, mappings and DMA pins.
> Keep eligible folios locked, flush pending writable TLB entries once,
> then page them out. Process a final partial batch at the end of the scan.
>
> Changes
> -------
> Patch 1: Extract the folio activation block at activate_locked into
> folio_activate_locked().
> Patch 2: Extract the folio-freeing path (buffer release, lazyfree,
> __remove_mapping, folio_batch drain) into folio_try_reclaim_free().
> Group retained folios, deferred frees and the reclaimed count
> in reclaim_folio_state, local to each shrink_folio_list() call.
> Patch 3: Extract the pageout() dispatch state machine into folio_try_pageout().
> Patch 4: Extract the TTU setup and try_to_unmap() block into folio_try_unmap().
> Patch 5: Implement batch TLB flushing logic. Dirty folios are accumulated in
> batches and a single TLB flush is performed for each batch before
> calling pageout.
>
> Testing
> -------
> The benchmark script uses stress-ng to compare TLB shootdown behavior before and
> after this patch. It constrains a stress-ng workload via memcg to force reclaim
> through shrink_folio_list(), reporting TLB shootdowns and IPIs.
>
> Core benchmark command: stress-ng --vm 16 --vm-bytes 2G --vm-keep --timeout 60
>
> ==========================================================================
> batch_dirty_tlb_flush Benchmark Results
> ==========================================================================
> Kernel: 7.0.0-rc1+ CPUs: 16
> MemTotal: 31834M SwapTotal: 8191M
> memcg limit: 512M alloc: 2G workers: 16 duration: 60s
> --------------------------------------------------------------------------
> Metric Before After Delta (abs / %)
> --------------------------------------------------------------------------
> bogo ops/s 28238.63 35833.97 +7595.34 (+26.9%)
> TLB shootdowns 55428953 17621697 -37807256 (-68.2%)
> Function call IPIs 34073695 14498768 -19574927 (-57.4%)
> pgscan_anon (pages) 52856224 60252894 7396670 (+14.0%)
> pgsteal_anon (pages) 29004962 34054753 5049791 (+17.4%)
> --------------------------------------------------------------------------
>
> Suggested-by: Kairui Song <kasong@xxxxxxxxxxx>
> Signed-off-by: Zhang Peng <bruzzhang@xxxxxxxxxxx>
> ---
Hi Peng,
Just a quick update: this patch is still on my radar. I ran a
quick perf test while building the kernel. Without your patch,
try_to_unmap_flush_dirty() accounted for 1.75% of CPU cycles.
I guess there could be workloads where try_to_unmap_flush_dirty()
accounts for an even larger share of CPU cycles.
# Children Self Command Shared Object Symbol
# ........ ........ ....... ................. ............................
#
1.75% 0.01% cc1 [kernel.kallsyms] [.] try_to_unmap_flush_dirty
|
--1.74%--try_to_unmap_flush_dirty
try_to_unmap_flush (inlined)
|
--1.73%--arch_tlbbatch_flush
|
--1.70%--flush_tlb_multi (inlined)
__flush_tlb_multi (inlined)
|
--1.69%--native_flush_tlb_multi
|
--1.68%--on_each_cpu_cond_mask
|
--1.62%--smp_call_function_many_cond
|
--1.23%--csd_lock_wait (inlined)
With your patch, the CPU overhead drops by roughly 7x:
# Children Self Command Shared Object Symbol
# ........ ........ ....... ................. ............................
#
0.25% 0.00% cc1 [kernel.kallsyms] [.] try_to_unmap_flush_dirty
|
--0.24%--try_to_unmap_flush_dirty
try_to_unmap_flush (inlined)
|
--0.24%--arch_tlbbatch_flush
|
--0.24%--flush_tlb_multi (inlined)
__flush_tlb_multi (inlined)
|
--0.24%--native_flush_tlb_multi
|
--0.23%--on_each_cpu_cond_mask
|
--0.21%--smp_call_function_many_cond
|
--0.14%--csd_lock_wait (inlined)
0.00% 0.00% fixdep [kernel.kallsyms] [.] try_to_unmap_flush_dirty
Overall, I think this is a good approach. I'll be traveling soon
and will return to this patchset for more testing and a thorough
review in about two and a half weeks.
Thanks
Barry