| Message ID | 20240429072417.2146732-5-wangkefeng.wang@huawei.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: filemap: try to batch lruvec stat updating | expand |
+ memcg maintainers and David too, please check all patches from link https://lore.kernel.org/linux-mm/20240429072417.2146732-1-wangkefeng.wang@huawei.com/ Thanks On 2024/4/29 15:24, Kefeng Wang wrote: > The filemap_map_pages() tries to map few pages(eg, 16 pages), but the > lruvec stat updating is called on each mapping, since the updating is > time-consuming, especially with memcg, so try to batch it when the memcg > and pgdat are same during the mapping, if luckily, we could save most of > time of lruvec stat updating, the lat_pagefault shows 3~4% improvement. > > Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com> > --- > mm/filemap.c | 33 ++++++++++++++++++++++++++++++--- > 1 file changed, 30 insertions(+), 3 deletions(-) > > diff --git a/mm/filemap.c b/mm/filemap.c > index 3966b6616d02..b27281707098 100644 > --- a/mm/filemap.c > +++ b/mm/filemap.c > @@ -3615,6 +3615,20 @@ static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf, > return ret; > } > > +static void filemap_lruvec_stat_update(struct mem_cgroup *memcg, > + pg_data_t *pgdat, int nr) > +{ > + struct lruvec *lruvec; > + > + if (!memcg) { > + __mod_node_page_state(pgdat, NR_FILE_MAPPED, nr); > + return; > + } > + > + lruvec = mem_cgroup_lruvec(memcg, pgdat); > + __mod_lruvec_state(lruvec, NR_FILE_MAPPED, nr); > +} > + > vm_fault_t filemap_map_pages(struct vm_fault *vmf, > pgoff_t start_pgoff, pgoff_t end_pgoff) > { > @@ -3628,6 +3642,9 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, > vm_fault_t ret = 0; > unsigned long rss = 0; > unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type; > + struct mem_cgroup *memcg, *memcg_cur; > + pg_data_t *pgdat, *pgdat_cur; > + int nr_mapped = 0; > > rcu_read_lock(); > folio = next_uptodate_folio(&xas, mapping, end_pgoff); > @@ -3648,9 +3665,20 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, > } > > folio_type = mm_counter_file(folio); > + memcg = folio_memcg(folio); > + pgdat = folio_pgdat(folio); > do { > unsigned long end; > - int nr_mapped = 0; > + > + memcg_cur = folio_memcg(folio); > + pgdat_cur = folio_pgdat(folio); > + > + if (unlikely(memcg != memcg_cur || pgdat != pgdat_cur)) { > + filemap_lruvec_stat_update(memcg, pgdat, nr_mapped); > + nr_mapped = 0; > + memcg = memcg_cur; > + pgdat = pgdat_cur; > + } > > addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT; > vmf->pte += xas.xa_index - last_pgoff; > @@ -3668,11 +3696,10 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, > nr_pages, &rss, &nr_mapped, > &mmap_miss); > > - __lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr_mapped); > - > folio_unlock(folio); > folio_put(folio); > } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL); > + filemap_lruvec_stat_update(memcg, pgdat, nr_mapped); > add_mm_counter(vma->vm_mm, folio_type, rss); > pte_unmap_unlock(vmf->pte, vmf->ptl); > out:
On Tue, May 07, 2024 at 05:06:57PM +0800, Kefeng Wang wrote: > > +static void filemap_lruvec_stat_update(struct mem_cgroup *memcg, > > + pg_data_t *pgdat, int nr) > > +{ > > + struct lruvec *lruvec; > > + > > + if (!memcg) { > > + __mod_node_page_state(pgdat, NR_FILE_MAPPED, nr); > > + return; > > + } > > + > > + lruvec = mem_cgroup_lruvec(memcg, pgdat); > > + __mod_lruvec_state(lruvec, NR_FILE_MAPPED, nr); > > +} > > + > > vm_fault_t filemap_map_pages(struct vm_fault *vmf, > > pgoff_t start_pgoff, pgoff_t end_pgoff) > > { > > @@ -3628,6 +3642,9 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, > > vm_fault_t ret = 0; > > unsigned long rss = 0; > > unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type; > > + struct mem_cgroup *memcg, *memcg_cur; > > + pg_data_t *pgdat, *pgdat_cur; > > + int nr_mapped = 0; > > > > rcu_read_lock(); > > folio = next_uptodate_folio(&xas, mapping, end_pgoff); > > @@ -3648,9 +3665,20 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, > > } > > > > folio_type = mm_counter_file(folio); > > + memcg = folio_memcg(folio); > > + pgdat = folio_pgdat(folio); You should be able to do: lruvec = folio_lruvec(folio); and then pass that directly to filemap_lruvec_stat_update().
On 2024/5/9 22:01, Johannes Weiner wrote: > On Tue, May 07, 2024 at 05:06:57PM +0800, Kefeng Wang wrote: >>> +static void filemap_lruvec_stat_update(struct mem_cgroup *memcg, >>> + pg_data_t *pgdat, int nr) >>> +{ >>> + struct lruvec *lruvec; >>> + >>> + if (!memcg) { >>> + __mod_node_page_state(pgdat, NR_FILE_MAPPED, nr); >>> + return; >>> + } >>> + >>> + lruvec = mem_cgroup_lruvec(memcg, pgdat); >>> + __mod_lruvec_state(lruvec, NR_FILE_MAPPED, nr); >>> +} >>> + >>> vm_fault_t filemap_map_pages(struct vm_fault *vmf, >>> pgoff_t start_pgoff, pgoff_t end_pgoff) >>> { >>> @@ -3628,6 +3642,9 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, >>> vm_fault_t ret = 0; >>> unsigned long rss = 0; >>> unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type; >>> + struct mem_cgroup *memcg, *memcg_cur; >>> + pg_data_t *pgdat, *pgdat_cur; >>> + int nr_mapped = 0; >>> >>> rcu_read_lock(); >>> folio = next_uptodate_folio(&xas, mapping, end_pgoff); >>> @@ -3648,9 +3665,20 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, >>> } >>> >>> folio_type = mm_counter_file(folio); >>> + memcg = folio_memcg(folio); >>> + pgdat = folio_pgdat(folio); > > You should be able to do: > > lruvec = folio_lruvec(folio); > > and then pass that directly to filemap_lruvec_stat_update(). It's obviously better, will update and address David's comment in patch3, thank you.
diff --git a/mm/filemap.c b/mm/filemap.c index 3966b6616d02..b27281707098 100644 --- a/mm/filemap.c +++ b/mm/filemap.c @@ -3615,6 +3615,20 @@ static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf, return ret; } +static void filemap_lruvec_stat_update(struct mem_cgroup *memcg, + pg_data_t *pgdat, int nr) +{ + struct lruvec *lruvec; + + if (!memcg) { + __mod_node_page_state(pgdat, NR_FILE_MAPPED, nr); + return; + } + + lruvec = mem_cgroup_lruvec(memcg, pgdat); + __mod_lruvec_state(lruvec, NR_FILE_MAPPED, nr); +} + vm_fault_t filemap_map_pages(struct vm_fault *vmf, pgoff_t start_pgoff, pgoff_t end_pgoff) { @@ -3628,6 +3642,9 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, vm_fault_t ret = 0; unsigned long rss = 0; unsigned int nr_pages = 0, mmap_miss = 0, mmap_miss_saved, folio_type; + struct mem_cgroup *memcg, *memcg_cur; + pg_data_t *pgdat, *pgdat_cur; + int nr_mapped = 0; rcu_read_lock(); folio = next_uptodate_folio(&xas, mapping, end_pgoff); @@ -3648,9 +3665,20 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, } folio_type = mm_counter_file(folio); + memcg = folio_memcg(folio); + pgdat = folio_pgdat(folio); do { unsigned long end; - int nr_mapped = 0; + + memcg_cur = folio_memcg(folio); + pgdat_cur = folio_pgdat(folio); + + if (unlikely(memcg != memcg_cur || pgdat != pgdat_cur)) { + filemap_lruvec_stat_update(memcg, pgdat, nr_mapped); + nr_mapped = 0; + memcg = memcg_cur; + pgdat = pgdat_cur; + } addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT; vmf->pte += xas.xa_index - last_pgoff; @@ -3668,11 +3696,10 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf, nr_pages, &rss, &nr_mapped, &mmap_miss); - __lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr_mapped); - folio_unlock(folio); folio_put(folio); } while ((folio = next_uptodate_folio(&xas, mapping, end_pgoff)) != NULL); + filemap_lruvec_stat_update(memcg, pgdat, nr_mapped); add_mm_counter(vma->vm_mm, folio_type, rss); pte_unmap_unlock(vmf->pte, vmf->ptl); out:
The filemap_map_pages() tries to map few pages(eg, 16 pages), but the lruvec stat updating is called on each mapping, since the updating is time-consuming, especially with memcg, so try to batch it when the memcg and pgdat are same during the mapping, if luckily, we could save most of time of lruvec stat updating, the lat_pagefault shows 3~4% improvement. Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com> --- mm/filemap.c | 33 ++++++++++++++++++++++++++++++--- 1 file changed, 30 insertions(+), 3 deletions(-)