| Message ID | 20240902072136.578720-1-alexs@kernel.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm/zsmalloc: add zpdesc memory descriptor for zswap.zpool | expand |
On (24/09/02 15:21), alexs@kernel.org wrote: > The descriptor still overlays the struct page; nothing has changed > in that regard. > [..] > This patachset does not increase the descriptor size nor introduce any > functional changes, and could save about 122Kbytes zsmalloc.o size. Is this number accurate? Where does such a large saving come from?
On (24/09/03 12:35), Sergey Senozhatsky wrote: > On (24/09/02 15:21), alexs@kernel.org wrote: > > The descriptor still overlays the struct page; nothing has changed > > in that regard. > > [..] > > This patachset does not increase the descriptor size nor introduce any > > functional changes, and could save about 122Kbytes zsmalloc.o size. > > Is this number accurate? Where does such a large saving come from? ./scripts/bloat-o-meter mm/zsmalloc.o-base mm/zsmalloc.o-patched add/remove: 0/0 grow/shrink: 2/6 up/down: 7/-843 (-836) Function old new delta obj_free 211 215 +4 zs_unmap_object 568 571 +3 zs_map_object 661 633 -28 SetZsPageMovable 135 86 -49 free_zspage 310 253 -57 __free_zspage 284 226 -58 zs_page_migrate 1304 1136 -168 async_free_zspage 1099 616 -483
On 9/3/24 11:45 AM, Sergey Senozhatsky wrote: > On (24/09/03 12:35), Sergey Senozhatsky wrote: >> On (24/09/02 15:21), alexs@kernel.org wrote: >>> The descriptor still overlays the struct page; nothing has changed >>> in that regard. >>> [..] >>> This patachset does not increase the descriptor size nor introduce any >>> functional changes, and could save about 122Kbytes zsmalloc.o size. >> >> Is this number accurate? Where does such a large saving come from? Hi Sergey, Thanks for asking, I collected some data on Aug 2, before the patchset, zsmalloc.o is 1200960 bytes, after patched, it's 1076144 bytes. At that time it's saved about 124KB, about 10%, with some debug option enabled. Just test again with x86defconfig + ZSMALLOC + ZSMALLOC_STAT, on mm-unstable tree, zsmalloc.o reduces to 34960 bytes from 37328 bytes, that's 6.3% saved. The bloat-o-meter show in machine: $ ./scripts/bloat-o-meter x86def/mm/zsmalloc.o x86def/mm/zsmalloc-patched.o add/remove: 1/1 grow/shrink: 1/5 up/down: 100/-1068 (-968) Function old new delta create_zpdesc_chain - 92 +92 zs_compact 2352 2360 +8 zs_malloc 1334 1277 -57 __free_zspage 291 223 -68 free_zspage 289 211 -78 create_page_chain 95 - -95 zs_page_migrate 950 807 -143 async_free_zspage 1252 625 -627 Total: Before=11868, After=10900, chg -8.16% > > ./scripts/bloat-o-meter mm/zsmalloc.o-base mm/zsmalloc.o-patched > add/remove: 0/0 grow/shrink: 2/6 up/down: 7/-843 (-836) > Function old new delta > obj_free 211 215 +4 > zs_unmap_object 568 571 +3 > zs_map_object 661 633 -28 > SetZsPageMovable 135 86 -49 > free_zspage 310 253 -57 > __free_zspage 284 226 -58 > zs_page_migrate 1304 1136 -168 > async_free_zspage 1099 616 -483 Many thanks for detailed data here! Thanks Alex
On 9/2/24 3:21 PM, alexs@kernel.org wrote: > From: Alex Shi <alexs@kernel.org> > ... > > This patchset abstracts the memory descriptor used in zsmalloc by zswap/zram. > The descriptor still overlays the struct page; nothing has changed > in that regard. What this patchset accomplishes is the use of folios in > to save some code size, and the introduction of a new concept, zpdesc. > This patchset is just an initial step; it does not bias the potential > changes to kmem_alloc or larger zspage modifications. > ... > > Thanks a lot for comments and suggestion from Yosry, Yoo, Sergey, Willy > and Vishal! > This patchset could save 6.3% code size, and it's a nice abstract of zsmalloc memory usage. Is there any more comments, or mind to give a reviewed-by? Thanks Alex
On Wed, Sep 04, 2024 at 02:54:14PM +0800, Alex Shi wrote: > > > On 9/2/24 3:21 PM, alexs@kernel.org wrote: > > From: Alex Shi <alexs@kernel.org> > > > ... > > > > > This patchset abstracts the memory descriptor used in zsmalloc by zswap/zram. > > The descriptor still overlays the struct page; nothing has changed > > in that regard. What this patchset accomplishes is the use of folios in > > to save some code size, and the introduction of a new concept, zpdesc. > > This patchset is just an initial step; it does not bias the potential > > changes to kmem_alloc or larger zspage modifications. > > > ... > > > > Thanks a lot for comments and suggestion from Yosry, Yoo, Sergey, Willy > > and Vishal! > > > > This patchset could save 6.3% code size, and it's a nice abstract of zsmalloc > memory usage. > Is there any more comments, or mind to give a reviewed-by? Please CC me on future versions. Most of the zsmalloc conversions seem ok, but I'd hold off on further iterations of the descriptor patches until the maintainers decide on what/how this descriptor will be used (i.e. our end goals). > Thanks > Alex
On 9/5/24 4:04 AM, Vishal Moola wrote: > On Wed, Sep 04, 2024 at 02:54:14PM +0800, Alex Shi wrote: >> >> >> On 9/2/24 3:21 PM, alexs@kernel.org wrote: >>> From: Alex Shi <alexs@kernel.org> >>> >> ... >> >>> >>> This patchset abstracts the memory descriptor used in zsmalloc by zswap/zram. >>> The descriptor still overlays the struct page; nothing has changed >>> in that regard. What this patchset accomplishes is the use of folios in >>> to save some code size, and the introduction of a new concept, zpdesc. >>> This patchset is just an initial step; it does not bias the potential >>> changes to kmem_alloc or larger zspage modifications. >>> >> ... >>> >>> Thanks a lot for comments and suggestion from Yosry, Yoo, Sergey, Willy >>> and Vishal! >>> >> >> This patchset could save 6.3% code size, and it's a nice abstract of zsmalloc >> memory usage. >> Is there any more comments, or mind to give a reviewed-by? > > Please CC me on future versions. Most of the zsmalloc conversions seem > ok, but I'd hold off on further iterations of the descriptor patches until > the maintainers decide on what/how this descriptor will be used > (i.e. our end goals). Thanks for your care. Is there some places or some conversion should be changed? btw, the descriptor is used now and could save a few code size. :) Thanks
From: Alex Shi <alexs@kernel.org> According to Metthew's plan, the page descriptor will be replace by a 8 bytes mem_desc on destination purpose. https://lore.kernel.org/lkml/YvV1KTyzZ+Jrtj9x@casper.infradead.org/ This patchset abstracts the memory descriptor used in zsmalloc by zswap/zram. The descriptor still overlays the struct page; nothing has changed in that regard. What this patchset accomplishes is the use of folios in to save some code size, and the introduction of a new concept, zpdesc. This patchset is just an initial step; it does not bias the potential changes to kmem_alloc or larger zspage modifications. To name the struct zpdesc instead of zsdesc, since there are still 3 zpools under zswap: zbud, z3fold, zsmalloc for now(z3fold maybe removed soon), and we could easyly extend it to other zswap.zpool in needs. For all zswap.zpools, they are all using single page since often used under memory pressure. So the conversion via folio series helper is better than page's for compound_head check saving. For now, all zpools are using some page struct members, like page.flags for PG_private/PG_locked. and list_head lru, page.mapping for page migration. This patachset does not increase the descriptor size nor introduce any functional changes, and could save about 122Kbytes zsmalloc.o size. Thanks a lot for comments and suggestion from Yosry, Yoo, Sergey, Willy and Vishal! Thanks Alex --- v6->v7: - rebase on akpm/mm-unstable on Aug 30 v5->v6: - rebase on akpm/mm-unstable on Aug 4 - recover 2 helpers and update code comments according to Vishal's comments on patch 1 and 6. v4->v5: - rebase on akpm/mm-unstable on Aug 8 + Stephen Rothwell's build fix. - add a helper and update code comments according to Sergey's comments - fold patch 20/21, remove 3 helpers functions according to Vishal's comments v3->v4: - rebase on akpm/mm-unstable Jul 21 - fixed a build warning reported by LKP - Add a comment update for struct page to zpdesc change v2->v3: - Fix LKP reported build issue - Update the Usage of struct zpdesc fields. - Rebase onto latest mm-unstable commit 2073cda629a4 v1->v2: - Take Yosry and Yoo's suggestion to add more members in zpdesc, - Rebase on latest mm-unstable commit 31334cf98dbd Alex Shi (10): mm/zsmalloc: add zpdesc memory descriptor for zswap.zpool mm/zsmalloc: use zpdesc in trylock_zspage()/lock_zspage() mm/zsmalloc: convert create_page_chain() and its users to use zpdesc mm/zsmalloc: rename reset_page to reset_zpdesc and use zpdesc in it mm/zsmalloc: convert SetZsPageMovable and remove unused funcs mm/zsmalloc: convert get/set_first_obj_offset() to take zpdesc mm/zsmalloc: introduce __zpdesc_clear_movable mm/zsmalloc: introduce __zpdesc_clear/set_zsmalloc() mm/zsmalloc: introduce zpdesc_clear_first() helper mm/zsmalloc: update comments for page->zpdesc changes Hyeonggon Yoo (11): mm/zsmalloc: convert __zs_map_object/__zs_unmap_object to use zpdesc mm/zsmalloc: add and use pfn/zpdesc seeking funcs mm/zsmalloc: convert obj_malloc() to use zpdesc mm/zsmalloc: convert obj_allocated() and related helpers to use zpdesc mm/zsmalloc: convert init_zspage() to use zpdesc mm/zsmalloc: convert obj_to_page() and zs_free() to use zpdesc mm/zsmalloc: add zpdesc_is_isolated()/zpdesc_zone() helper for zs_page_migrate() mm/zsmalloc: convert __free_zspage() to use zdsesc mm/zsmalloc: convert location_to_obj() to take zpdesc mm/zsmalloc: convert migrate_zspage() to use zpdesc mm/zsmalloc: convert get_zspage() to take zpdesc mm/zpdesc.h | 147 +++++++++++++++ mm/zsmalloc.c | 490 +++++++++++++++++++++++++++----------------------- 2 files changed, 414 insertions(+), 223 deletions(-) create mode 100644 mm/zpdesc.h