[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <345161ac-3b42-48aa-ab3d-3b183316479a@redhat.com>
Date: Fri, 31 May 2024 16:32:04 +0200
From: David Hildenbrand <david@...hat.com>
To: Matthew Wilcox <willy@...radead.org>,
Sergey Senozhatsky <senozhatsky@...omium.org>
Cc: linux-kernel@...r.kernel.org, linux-mm@...ck.org,
Andrew Morton <akpm@...ux-foundation.org>, Mike Rapoport <rppt@...nel.org>,
Minchan Kim <minchan@...nel.org>, Hyeonggon Yoo <42.hyeyoo@...il.com>
Subject: Re: [PATCH v2 3/6] mm/zsmalloc: use a proper page type
On 31.05.24 16:27, Matthew Wilcox wrote:
> On Thu, May 30, 2024 at 02:01:23PM +0900, Sergey Senozhatsky wrote:
>> On (24/05/29 13:19), David Hildenbrand wrote:
>>> We won't be able to support 256 KiB base pages, which is acceptable.
>> [..]
>>> +config HAVE_ZSMALLOC
>>> + def_bool y
>>> + depends on MMU
>>> + depends on PAGE_SIZE_LESS_THAN_256KB # we want <= 64 KiB
>>
>> Can't really say that I'm happy with this, but if mm-folks are
>> fine then okay.
>
> I have an idea ...
>
> We use 6 of the bits in the top byte of the page_type to enumerate
> a type (ie value 0x80-0xbf) and then the remaining 24 bits are
> available. It's actually more efficient:
>
> $ ./scripts/bloat-o-meter prev.o .build-debian/mm/filemap.o
> add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-40 (-40)
> Function old new delta
> __filemap_add_folio 1102 1098 -4
> filemap_unaccount_folio 455 446 -9
> replace_page_cache_folio 474 447 -27
> Total: Before=41258, After=41218, chg -0.10%
>
> (that's all from PG_hugetlb)
>
> before:
> 1406: 8b 46 30 mov 0x30(%rsi),%eax
> mapcount = atomic_read(&folio->_mapcount) + 1;
> 1409: 83 c0 01 add $0x1,%eax
> if (mapcount < PAGE_MAPCOUNT_RESERVE + 1)
> 140c: 83 f8 81 cmp $0xffffff81,%eax
> 140f: 7d 6c jge 147d <filemap_unaccount_folio+0x8d>
> 1411: 8b 43 30 mov 0x30(%rbx),%eax
> 1414: 25 00 08 00 f0 and $0xf0000800,%eax
> 1419: 3d 00 00 00 f0 cmp $0xf0000000,%eax
> 141e: 74 4e je 146e <filemap_unaccount_folio+0x7e>
>
> after:
> 1406: 8b 46 30 mov 0x30(%rsi),%eax
> mapcount = atomic_read(&folio->_mapcount) + 1;
> 1409: 83 c0 01 add $0x1,%eax
> if (mapcount < PAGE_MAPCOUNT_RESERVE + 1)
> 140c: 83 f8 81 cmp $0xffffff81,%eax
> 140f: 7d 63 jge 1474 <filemap_unaccount_folio+0x8
> 4>
> if (folio_test_hugetlb(folio))
> 1411: 80 7b 33 84 cmpb $0x84,0x33(%rbx)
> 1415: 74 4e je 1465 <filemap_unaccount_folio+0x75>
>
> so we go from "mov, and, cmp, je" to just "cmpb, je", which must surely
> be faster to execute as well as being more compact in the I$ (6 bytes vs 15).
>
> Anyway, not tested but this is the patch I used to generate the above.
> More for comment than application.
Right, it's likely very similar to my previous proposal to use 8 bit
(uint8_t) for the type.
https://lore.kernel.org/all/00ba1dff-7c05-46e8-b0d9-a78ac1cfc198@redhat.com/
I would prefer if we would do that separately; unless someone is able to
raise why we care about zram + 256KiB that much right now. (claim: we don't)
>
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index 5265b3434b9e..4129d04ac812 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -942,24 +942,24 @@ PAGEFLAG_FALSE(HasHWPoisoned, has_hwpoisoned)
> * mistaken for a page type value.
> */
>
> -#define PAGE_TYPE_BASE 0xf0000000
> -/* Reserve 0x0000007f to catch underflows of _mapcount */
> -#define PAGE_MAPCOUNT_RESERVE -128
> -#define PG_buddy 0x00000080
> -#define PG_offline 0x00000100
> -#define PG_table 0x00000200
> -#define PG_guard 0x00000400
> -#define PG_hugetlb 0x00000800
> -#define PG_slab 0x00001000
> -
> -#define PageType(page, flag) \
> - ((page->page_type & (PAGE_TYPE_BASE | flag)) == PAGE_TYPE_BASE)
> -#define folio_test_type(folio, flag) \
> - ((folio->page.page_type & (PAGE_TYPE_BASE | flag)) == PAGE_TYPE_BASE)
> +/* Reserve 0x0000007f to catch underflows of _mapcount */
> +#define PAGE_MAPCOUNT_RESERVE -128
> +
> +#define PG_buddy 0x80
> +#define PG_offline 0x81
> +#define PG_table 0x82
> +#define PG_guard 0x83
> +#define PG_hugetlb 0x84
> +#define PG_slab 0x85
Hoping we can stop calling that PG_ ...
--
Cheers,
David / dhildenb
Powered by blists - more mailing lists