On Sat, Feb 17, 2024 at 9:52 PM Chengming Zhou <chengming.zhou@xxxxxxxxx> wrote: > > On 2024/2/15 15:06, Yu Zhao wrote: > > On Wed, Feb 14, 2024 at 4:18 AM Chengming Zhou <chengming.zhou@xxxxxxxxx> wrote: > >> > >> On 2024/2/14 15:13, Yu Zhao wrote: > >>> On Fri, Feb 9, 2024 at 6:00 AM <chengming.zhou@xxxxxxxxx> wrote: > >>>> > >>>> From: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx> > >>>> > >>>> All LRU move interfaces have a problem that it has no effect if the > >>>> folio is isolated from LRU (in cpu batch or isolated by shrinker). > >>>> Since it can't move/change folio LRU status when it's isolated, mostly > >>>> just clear the folio flag and do nothing in this case. > >>>> > >>>> In our case, a written back and reclaimable folio won't be rotated to > >>>> the tail of inactive list, since it's still in cpu lru_add batch. It > >>>> may cause the delayed reclaim of this folio and evict other folios. > >>>> > >>>> This patch changes to queue the reclaimable folio to cpu rotate batch > >>>> even when !folio_test_lru(), hoping it will likely be handled after > >>>> the lru_add batch which will put folio on the LRU list first, so > >>>> will be rotated to the tail successfully when handle rotate batch. > >>>> > >>>> Signed-off-by: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx> > >>> > >>> I don't think the analysis is correct. IIRC, writeback from non > >>> reclaim paths doesn't require isolation and the reclaim path doesn't > >>> use struct folio_batch lru_add. > >> > >> Ah, my bad, I forgot to mention the important context in the message: > >> > >> This is not from the normal reclaim context, it's from zswap writeback > >> reclaim context, which will first set PG_reclaim flag, then submit the > >> async writeback io. > >> > >> If the writeback io complete fast enough, folio_rotate_reclaimable() > >> will be called before that folio put on LRU list (it still in the local > >> lru_add batch, so it's somewhat like isolated too) > >> > >>> > >>> Did you see any performance improvements with this patch? In general, > >>> this kind of patches should have performance numbers to show it really > >>> helps (not just in theory). > >> > >> Right, there are some improvements, the numbers are put in cover letter. > >> But this solution is not good enough, just RFC for discussion. :) > >> > >> mm-unstable-hot zswap-lru-reclaim > >> real 63.34 62.72 > >> user 1063.20 1060.30 > >> sys 272.04 256.14 > >> workingset_refault_anon 2103297.00 1788155.80 > >> workingset_refault_file 28638.20 39249.40 > >> workingset_activate_anon 746134.00 695435.40 > >> workingset_activate_file 4344.60 4255.80 > >> workingset_restore_anon 653163.80 605315.60 > >> workingset_restore_file 1079.00 883.00 > >> workingset_nodereclaim 0.00 0.00 > >> pgscan 12971305.60 12730331.20 > >> pgscan_kswapd 0.00 0.00 > >> pgscan_direct 12971305.60 12730331.20 > >> pgscan_khugepaged 0.00 0.00 > >> > >>> > >>> My guess is that you are hitting this problem [1]. > >>> > >>> [1] https://lore.kernel.org/linux-mm/20221116013808.3995280-1-yuzhao@xxxxxxxxxx/ > >> > >> Right, I just see it, it's the same problem. The only difference is that > >> in your case the folio is isolated by shrinker, in my case, the folio is > >> in cpu lru_add batch. Anyway, the result is the same, that folio can't be > >> rotated successfully when writeback complete. > > > > In that case, a better solution would be to make lru_add add > > (_reclaim() && !_dirty() && !_writeback()) folios at the tail. > > (_rotate() needs to leave _reclaim() set if it fails to rotate.) > > Right, this is a solution. But PG_readahead is alias of PG_reclaim, > I'm afraid this would rotate readahead folio to the inactive tail. Then drain before setting readahead, since readahead isn't set on every folio.
