Re: [PATCH v2 1/2] mm:vmscan: the dirty folio in folio_list skip unmap

[zhiguojiang]

在 2023/10/20 12:15, Matthew Wilcox 写道:
> On Fri, Oct 20, 2023 at 11:59:33AM +0800, zhiguojiang wrote:
> > > > @@ -1261,43 +1305,6 @@ static unsigned int shrink_folio_list(struct
> > > > list_head *folio_list,
> > > >                        enum ttu_flags flags = TTU_BATCH_FLUSH;
> > > >                        bool was_swapbacked =
> > > > folio_test_swapbacked(folio);
> > > >
> > > > -                     if (folio_test_dirty(folio)) {
> > > > -                             /*
> > > > -                              * Only kswapd can writeback
> > > > filesystem folios
> > > > -                              * to avoid risk of stack overflow.
> > > > But avoid
> > > > -                              * injecting inefficient single-folio
> > > > I/O into
> > > > -                              * flusher writeback as much as
> > > > possible: only
> > > > -                              * write folios when we've encountered
> > > > many
> > > > -                              * dirty folios, and when we've
> > > > already scanned
> > > > -                              * the rest of the LRU for clean
> > > > folios and see
> > > > -                              * the same dirty folios again (with
> > > > the reclaim
> > > > -                              * flag set).
> > > > -                              */
> > > > -                             if (folio_is_file_lru(folio) &&
> > > > -                                     (!current_is_kswapd() ||
> > > > - !folio_test_reclaim(folio) ||
> > > > -                                      !test_bit(PGDAT_DIRTY,
> > > > &pgdat->flags))) {
> > > > -                                     /*
> > > > -                                      * Immediately reclaim when
> > > > written back.
> > > > -                                      * Similar in principle to
> > > > folio_deactivate()
> > > > -                                      * except we already have the
> > > > folio isolated
> > > > -                                      * and know it's dirty
> > > > -                                      */
> > > > -                                     node_stat_mod_folio(folio,
> > > > NR_VMSCAN_IMMEDIATE,
> > > > -                                                     nr_pages);
> > > > -                                     folio_set_reclaim(folio);
> > > > -
> > > > -                                     goto activate_locked;
> > > > -                             }
> > > > -
> > > > -                             if (references == FOLIOREF_RECLAIM_CLEAN)
> > > > -                                     goto keep_locked;
> > > > -                             if (!may_enter_fs(folio, sc->gfp_mask))
> > > > -                                     goto keep_locked;
> > > > -                             if (!sc->may_writepage)
> > > > -                                     goto keep_locked;
> > > > -                     }
> > > > -
> > > >                        if (folio_test_pmd_mappable(folio))
> > > >                                flags |= TTU_SPLIT_HUGE_PMD;
> > > I'm confused. Did you apply this on top of v1 by accident?
> > Hi,
> > According to my modified mm_vmscan_lru_shrink_inactive test tracelog, in the
> You're missing David's point.  You've generated this patch against ...
> something ... that isn't upstream.  Probably against v1 of your
> patch.  Please check your git tree.
>
> > 32 scanned inactive file pages, 20 were dirty, and the 20 dirty pages were
> > not reclamed, but they took 20us to perform try_to_unmap.
> >
> > I think unreclaimed dirty folio in inactive file lru can skip to perform
> > try_to_unmap. Please help to continue review. Thanks.
> >
> > kswapd0-99      (     99) [005] .....   687.793724:
> > mm_vmscan_lru_shrink_inactive: [Justin] nid 0 scan=32 isolate=32 reclamed=12
> > nr_dirty=20 nr_unqueued_dirty=20 nr_writeback=0 nr_congested=0
> > nr_immediate=0 nr_activate[0]=0 nr_activate[1]=20 nr_ref_keep=0
> > nr_unmap_fail=0 priority=2 file=RECLAIM_WB_FILE|RECLAIM_WB_ASYNC total=39
> > exe=0 reference_cost=5 reference_exe=0 unmap_cost=21 unmap_exe=0
> > dirty_unmap_cost=20 dirty_unmap_exe=0 pageout_cost=0 pageout_exe=0
> Are you seeing measurable changes for any workloads?  It certainly seems
> like you should, but it would help if you chose a test from mmtests and
> showed how performance changed on your system.
In one mmtest, the max times for a invalid recyling of a folio_list 
dirty folio that does not support pageout and has been activated in 
shrink_folio_list() are: cost=51us, exe=2365us.

Calculate according to this formula: dirty_cost / total_cost * 100%, the 
recyling efficiency of dirty folios can be improved 53.13%、82.95%.

So this patch can optimize shrink efficiency and reduce the workload of 
kswapd to a certain extent.

kswapd0-96      (     96) [005] .....   387.218548: 
mm_vmscan_lru_shrink_inactive: [Justin] nid 0 nr_scanned 32 nr_taken 32 
nr_reclaimed 31 nr_dirty  1 nr_unqueued_dirty  1 nr_writeback 0 
nr_activate[1]  1 nr_ref_keep  0 f RECLAIM_WB_FILE|RECLAIM_WB_ASYNC 
total_cost 96 total_exe 2365 dirty_cost 51 total_exe 2365

kswapd0-96      (     96) [006] .....   412.822532: 
mm_vmscan_lru_shrink_inactive: [Justin] nid 0 nr_scanned 32 nr_taken 32 
nr_reclaimed  0 nr_dirty 32 nr_unqueued_dirty 32 nr_writeback 0 
nr_activate[1] 19 nr_ref_keep 13 f RECLAIM_WB_FILE|RECLAIM_WB_ASYNC 
total_cost 88 total_exe 605  dirty_cost 73 total_exe 605