Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Chuck Lever III <chuck.lever@oracle.com>
To:     Mel Gorman <mgorman@techsingularity.net>
CC:     Jesper Dangaard Brouer <brouer@redhat.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Vlastimil Babka <vbabka@suse.cz>,
        Christoph Hellwig <hch@infradead.org>,
        Alexander Duyck <alexander.duyck@gmail.com>,
        Matthew Wilcox <willy@infradead.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Linux-Net <netdev@vger.kernel.org>,
        Linux-MM <linux-mm@kvack.org>,
        Linux NFS Mailing List <linux-nfs@vger.kernel.org>
Subject: Re: [PATCH 0/3 v5] Introduce a bulk order-0 page allocator
Thread-Topic: [PATCH 0/3 v5] Introduce a bulk order-0 page allocator
Thread-Index: AQHXHvxh39vYR/ByH0SHQRVYSjGN/KqQU18AgAAXrwCAAAwKgIAAByWAgADtmYCAADyVgIAARPsA
Date:   Tue, 23 Mar 2021 18:52:34 +0000
Message-ID: <AD267230-8422-4AE4-A315-82C8DE00AA46@oracle.com>
References: <20210322091845.16437-1-mgorman@techsingularity.net>
 <C1DEE677-47B2-4B12-BA70-6E29F0D199D9@oracle.com>
 <20210322194948.GI3697@techsingularity.net>
 <0E0B33DE-9413-4849-8E78-06B0CDF2D503@oracle.com>
 <20210322205827.GJ3697@techsingularity.net> <20210323120851.18d430cf@carbon>
 <20210323144541.GL3697@techsingularity.net>
In-Reply-To: <20210323144541.GL3697@techsingularity.net>
Accept-Language: en-US
Content-Language: en-US
Content-Type: text/plain; charset="us-ascii"
Content-ID: <41A98A0644136B48A8B6546D31C06902@namprd10.prod.outlook.com>
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0
X-MS-Exchange-CrossTenant-Network-Message-Id: 89b85105-2c6a-4e97-0eac-08d8ee2cd277
X-MS-Exchange-CrossTenant-originalarrivaltime: 23 Mar 2021 18:52:34.9436
 (UTC)
X-MS-Exchange-CrossTenant-fromentityheader: Hosted
X-MS-Exchange-CrossTenant-id: 4e2c6054-71cb-48f1-bd6c-3a9705aca71b
X-MS-Exchange-CrossTenant-mailboxtype: HOSTED
X-MS-Exchange-CrossTenant-userprincipalname: y757Trml1rqqZwNyOxLkMvDZ0MvRimDJmQzfzB3ApBt7OICo5eaoL4UlBxVij9T4E/dbWfNW6fECg1fRQ04odw==
X-MS-Exchange-Transport-CrossTenantHeadersStamped: BYAPR10MB3110
X-Proofpoint-Virus-Version: vendor=nai engine=6200 definitions=9932 signatures=668683
X-Proofpoint-Spam-Details: rule=notspam policy=default score=0 adultscore=0 malwarescore=0 spamscore=0
 mlxscore=0 phishscore=0 suspectscore=0 mlxlogscore=999 bulkscore=0
 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.12.0-2009150000
 definitions=main-2103230136
X-Proofpoint-Virus-Version: vendor=nai engine=6200 definitions=9932 signatures=668683
X-Proofpoint-Spam-Details: rule=notspam policy=default score=0 adultscore=0 malwarescore=0 mlxscore=0
 priorityscore=1501 bulkscore=0 impostorscore=0 lowpriorityscore=0
 phishscore=0 mlxlogscore=999 suspectscore=0 clxscore=1015 spamscore=0
 classifier=spam adjust=0 reason=mlx scancount=1 engine=8.12.0-2009150000
 definitions=main-2103230136
Precedence: bulk
List-ID: <linux-nfs.vger.kernel.org>
X-Mailing-List: linux-nfs@vger.kernel.org


> On Mar 23, 2021, at 10:45 AM, Mel Gorman <mgorman@techsingularity.net> wr=
ote:
>=20
> On Tue, Mar 23, 2021 at 12:08:51PM +0100, Jesper Dangaard Brouer wrote:
>>>> <SNIP>
>>>> My results show that, because svc_alloc_arg() ends up calling
>>>> __alloc_pages_bulk() twice in this case, it ends up being
>>>> twice as expensive as the list case, on average, for the same
>>>> workload.
>>>>=20
>>>=20
>>> Ok, so in this case the caller knows that holes are always at the
>>> start. If the API returns an index that is a valid index and populated,
>>> it can check the next index and if it is valid then the whole array
>>> must be populated.
>>>=20
>>> <SNIP>
>>=20
>> I do know that I suggested moving prep_new_page() out of the
>> IRQ-disabled loop, but maybe was a bad idea, for several reasons.
>>=20
>> All prep_new_page does is to write into struct page, unless some
>> debugging stuff (like kasan) is enabled. This cache-line is hot as
>> LRU-list update just wrote into this cache-line.  As the bulk size goes
>> up, as Matthew pointed out, this cache-line might be pushed into
>> L2-cache, and then need to be accessed again when prep_new_page() is
>> called.
>>=20
>> Another observation is that moving prep_new_page() into loop reduced
>> function size with 253 bytes (which affect I-cache).
>>=20
>>   ./scripts/bloat-o-meter mm/page_alloc.o-prep_new_page-outside mm/page_=
alloc.o-prep_new_page-inside
>>    add/remove: 18/18 grow/shrink: 0/1 up/down: 144/-397 (-253)
>>    Function                                     old     new   delta
>>    __alloc_pages_bulk                          1965    1712    -253
>>    Total: Before=3D60799, After=3D60546, chg -0.42%
>>=20
>> Maybe it is better to keep prep_new_page() inside the loop.  This also
>> allows list vs array variant to share the call.  And it should simplify
>> the array variant code.
>>=20
>=20
> I agree. I did not like the level of complexity it incurred for arrays
> or the fact it required that a list to be empty when alloc_pages_bulk()
> is called. I thought the concern for calling prep_new_page() with IRQs
> disabled was a little overblown but did not feel strongly enough to push
> back on it hard given that we've had problems with IRQs being disabled
> for long periods before. At worst, at some point in the future we'll have
> to cap the number of pages that can be requested or enable/disable IRQs
> every X pages.
>=20
> New candidate
>=20
> git://git.kernel.org/pub/scm/linux/kernel/git/mel/linux.git mm-bulk-rebas=
e-v6r4

I have rebased the SUNRPC patches to v6r4. Testing has shown a
minor functional regression, which I'm chasing down. But
performance is in the same ballpark. FYI


> Interface is still the same so a rebase should be trivial. Diff between
> v6r2 and v6r4 is as follows. I like the diffstat if nothing else :P
>=20
>=20
> mm/page_alloc.c | 54 +++++++++++++---------------------------------------=
--
> 1 file changed, 13 insertions(+), 41 deletions(-)
>=20
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 547a84f11310..be1e33a4df39 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -4999,25 +4999,20 @@ int __alloc_pages_bulk(gfp_t gfp, int preferred_n=
id,
> 	struct alloc_context ac;
> 	gfp_t alloc_gfp;
> 	unsigned int alloc_flags;
> -	int nr_populated =3D 0, prep_index =3D 0;
> -	bool hole =3D false;
> +	int nr_populated =3D 0;
>=20
> 	if (WARN_ON_ONCE(nr_pages <=3D 0))
> 		return 0;
>=20
> -	if (WARN_ON_ONCE(page_list && !list_empty(page_list)))
> -		return 0;
> -
> -	/* Skip populated array elements. */
> -	if (page_array) {
> -		while (nr_populated < nr_pages && page_array[nr_populated])
> -			nr_populated++;
> -		if (nr_populated =3D=3D nr_pages)
> -			return nr_populated;
> -		prep_index =3D nr_populated;
> -	}
> +	/*
> +	 * Skip populated array elements to determine if any pages need
> +	 * to be allocated before disabling IRQs.
> +	 */
> +	while (page_array && page_array[nr_populated] && nr_populated < nr_page=
s)
> +		nr_populated++;
>=20
> -	if (nr_pages =3D=3D 1)
> +	/* Use the single page allocator for one page. */
> +	if (nr_pages - nr_populated =3D=3D 1)
> 		goto failed;
>=20
> 	/* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
> @@ -5056,22 +5051,17 @@ int __alloc_pages_bulk(gfp_t gfp, int preferred_n=
id,
> 	if (!zone)
> 		goto failed;
>=20
> -retry_hole:
> 	/* Attempt the batch allocation */
> 	local_irq_save(flags);
> 	pcp =3D &this_cpu_ptr(zone->pageset)->pcp;
> 	pcp_list =3D &pcp->lists[ac.migratetype];
>=20
> 	while (nr_populated < nr_pages) {
> -		/*
> -		 * Stop allocating if the next index has a populated
> -		 * page or the page will be prepared a second time when
> -		 * IRQs are enabled.
> -		 */
> +
> +		/* Skip existing pages */
> 		if (page_array && page_array[nr_populated]) {
> -			hole =3D true;
> 			nr_populated++;
> -			break;
> +			continue;
> 		}
>=20
> 		page =3D __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
> @@ -5092,6 +5082,7 @@ int __alloc_pages_bulk(gfp_t gfp, int preferred_nid=
,
> 		__count_zid_vm_events(PGALLOC, zone_idx(zone), 1);
> 		zone_statistics(ac.preferred_zoneref->zone, zone);
>=20
> +		prep_new_page(page, 0, gfp, 0);
> 		if (page_list)
> 			list_add(&page->lru, page_list);
> 		else
> @@ -5101,25 +5092,6 @@ int __alloc_pages_bulk(gfp_t gfp, int preferred_ni=
d,
>=20
> 	local_irq_restore(flags);
>=20
> -	/* Prep pages with IRQs enabled. */
> -	if (page_list) {
> -		list_for_each_entry(page, page_list, lru)
> -			prep_new_page(page, 0, gfp, 0);
> -	} else {
> -		while (prep_index < nr_populated)
> -			prep_new_page(page_array[prep_index++], 0, gfp, 0);
> -
> -		/*
> -		 * If the array is sparse, check whether the array is
> -		 * now fully populated. Continue allocations if
> -		 * necessary.
> -		 */
> -		while (nr_populated < nr_pages && page_array[nr_populated])
> -			nr_populated++;
> -		if (hole && nr_populated < nr_pages)
> -			goto retry_hole;
> -	}
> -
> 	return nr_populated;
>=20
> failed_irq:
>=20
> --=20
> Mel Gorman
> SUSE Labs

--
Chuck Lever