Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
IronPort-SDR: JHhJeyumR54D6d6W4dACdt6aoabpDPgy2GcpLvbF4Yl/Tc5UOxO2qiBIPnpLCD86saw+HGeqNX
 GLihHr9kp4GbkACYdeBNXQwbeaYRRWeqdXu0k18XkcsUZwWAyZvQk76xj4omc3FgodTOYlkvCI
 nO2HHEMOLcvv16yqH+dH0TQpk8/5/72oSXrZNxgJsRJFn15Qtcj9C5IM0ZSJxER/5WGW3zJoNp
 PwntL8B2iwuWKFMQ6n4PPoOcpHEeRlyLwFm1SSNN1lJzWxeRCzKX8TPfFRn9wU8qVRvX5l3MDB
 SPw=
From:   Damien Le Moal <Damien.LeMoal@wdc.com>
To:     Ming Lei <ming.lei@redhat.com>
CC:     Changheun Lee <nanich.lee@samsung.com>,
        Johannes Thumshirn <Johannes.Thumshirn@wdc.com>,
        "asml.silence@gmail.com" <asml.silence@gmail.com>,
        "axboe@kernel.dk" <axboe@kernel.dk>,
        "linux-block@vger.kernel.org" <linux-block@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "osandov@fb.com" <osandov@fb.com>,
        "patchwork-bot@kernel.org" <patchwork-bot@kernel.org>,
        "tj@kernel.org" <tj@kernel.org>,
        "tom.leiming@gmail.com" <tom.leiming@gmail.com>,
        "jisoo2146.oh@samsung.com" <jisoo2146.oh@samsung.com>,
        "junho89.kim@samsung.com" <junho89.kim@samsung.com>,
        "mj0123.lee@samsung.com" <mj0123.lee@samsung.com>,
        "seunghwan.hyun@samsung.com" <seunghwan.hyun@samsung.com>,
        "sookwan7.kim@samsung.com" <sookwan7.kim@samsung.com>,
        "woosung2.lee@samsung.com" <woosung2.lee@samsung.com>,
        "yt0928.kim@samsung.com" <yt0928.kim@samsung.com>
Subject: Re: [PATCH v3 1/2] bio: limit bio max size
Thread-Topic: [PATCH v3 1/2] bio: limit bio max size
Thread-Index: AQHW84VRloabNXvGbkml+gtfdHfb2g==
Date:   Tue, 26 Jan 2021 06:26:02 +0000
Message-ID: <BL0PR04MB65144B93443C4F8EF9C48024E7BC9@BL0PR04MB6514.namprd04.prod.outlook.com>
References: <CGME20210126014805epcas1p2c4fc40f01c9c89b0a94ff6cac5408347@epcas1p2.samsung.com>
 <20210126013235.28711-1-nanich.lee@samsung.com>
 <20210126035748.GA1071341@T590>
 <BL0PR04MB65143B60B6062996764423C3E7BC9@BL0PR04MB6514.namprd04.prod.outlook.com>
 <20210126060724.GA1086419@T590>
Accept-Language: en-US
Content-Language: en-US
wdcipoutbound: EOP-TRUE
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: quoted-printable
MIME-Version: 1.0
X-MS-Exchange-CrossTenant-Network-Message-Id: dad677c2-3f02-40a3-3a46-08d8c1c3411d
X-MS-Exchange-CrossTenant-originalarrivaltime: 26 Jan 2021 06:26:02.7047
 (UTC)
X-MS-Exchange-CrossTenant-fromentityheader: Hosted
X-MS-Exchange-CrossTenant-id: b61c8803-16f3-4c35-9b17-6f65f441df86
X-MS-Exchange-CrossTenant-mailboxtype: HOSTED
X-MS-Exchange-CrossTenant-userprincipalname: JdiigEeanmYJwNWTLRvPVuMghuoQee2yXgNmZRiAalxgSiEkpaAu+5fU9HNhTfwiHAX/yi+1ZnntdoVmstuBHA==
X-MS-Exchange-Transport-CrossTenantHeadersStamped: BL0PR04MB4852
Precedence: bulk
List-ID: <linux-kernel.vger.kernel.org>
X-Mailing-List: linux-kernel@vger.kernel.org

On 2021/01/26 15:07, Ming Lei wrote:=0A=
> On Tue, Jan 26, 2021 at 04:06:06AM +0000, Damien Le Moal wrote:=0A=
>> On 2021/01/26 12:58, Ming Lei wrote:=0A=
>>> On Tue, Jan 26, 2021 at 10:32:34AM +0900, Changheun Lee wrote:=0A=
>>>> bio size can grow up to 4GB when muli-page bvec is enabled.=0A=
>>>> but sometimes it would lead to inefficient behaviors.=0A=
>>>> in case of large chunk direct I/O, - 32MB chunk read in user space -=
=0A=
>>>> all pages for 32MB would be merged to a bio structure if the pages=0A=
>>>> physical addresses are contiguous. it makes some delay to submit=0A=
>>>> until merge complete. bio max size should be limited to a proper size.=
=0A=
>>>>=0A=
>>>> When 32MB chunk read with direct I/O option is coming from userspace,=
=0A=
>>>> kernel behavior is below now. it's timeline.=0A=
>>>>=0A=
>>>>  | bio merge for 32MB. total 8,192 pages are merged.=0A=
>>>>  | total elapsed time is over 2ms.=0A=
>>>>  |------------------ ... ----------------------->|=0A=
>>>>                                                  | 8,192 pages merged =
a bio.=0A=
>>>>                                                  | at this time, first=
 bio submit is done.=0A=
>>>>                                                  | 1 bio is split to 3=
2 read request and issue.=0A=
>>>>                                                  |--------------->=0A=
>>>>                                                   |--------------->=0A=
>>>>                                                    |--------------->=
=0A=
>>>>                                                               ......=
=0A=
>>>>                                                                    |--=
------------->=0A=
>>>>                                                                     |-=
-------------->|=0A=
>>>>                           total 19ms elapsed to complete 32MB read don=
e from device. |=0A=
>>>>=0A=
>>>> If bio max size is limited with 1MB, behavior is changed below.=0A=
>>>>=0A=
>>>>  | bio merge for 1MB. 256 pages are merged for each bio.=0A=
>>>>  | total 32 bio will be made.=0A=
>>>>  | total elapsed time is over 2ms. it's same.=0A=
>>>>  | but, first bio submit timing is fast. about 100us.=0A=
>>>>  |--->|--->|--->|---> ... -->|--->|--->|--->|--->|=0A=
>>>>       | 256 pages merged a bio.=0A=
>>>>       | at this time, first bio submit is done.=0A=
>>>>       | and 1 read request is issued for 1 bio.=0A=
>>>>       |--------------->=0A=
>>>>            |--------------->=0A=
>>>>                 |--------------->=0A=
>>>>                                       ......=0A=
>>>>                                                  |--------------->=0A=
>>>>                                                   |--------------->|=
=0A=
>>>>         total 17ms elapsed to complete 32MB read done from device. |=
=0A=
>>>>=0A=
>>>> As a result, read request issue timing is faster if bio max size is li=
mited.=0A=
>>>> Current kernel behavior with multipage bvec, super large bio can be cr=
eated.=0A=
>>>> And it lead to delay first I/O request issue.=0A=
>>>>=0A=
>>>> Signed-off-by: Changheun Lee <nanich.lee@samsung.com>=0A=
>>>> ---=0A=
>>>>  block/bio.c            | 17 ++++++++++++++++-=0A=
>>>>  include/linux/bio.h    |  4 +++-=0A=
>>>>  include/linux/blkdev.h |  3 +++=0A=
>>>>  3 files changed, 22 insertions(+), 2 deletions(-)=0A=
>>>>=0A=
>>>> diff --git a/block/bio.c b/block/bio.c=0A=
>>>> index 1f2cc1fbe283..ec0281889045 100644=0A=
>>>> --- a/block/bio.c=0A=
>>>> +++ b/block/bio.c=0A=
>>>> @@ -287,6 +287,21 @@ void bio_init(struct bio *bio, struct bio_vec *ta=
ble,=0A=
>>>>  }=0A=
>>>>  EXPORT_SYMBOL(bio_init);=0A=
>>>>  =0A=
>>>> +unsigned int bio_max_size(struct bio *bio)=0A=
>>>> +{=0A=
>>>> +	struct request_queue *q;=0A=
>>>> +=0A=
>>>> +	if (!bio->bi_disk)=0A=
>>>> +		return UINT_MAX;=0A=
>>>> +=0A=
>>>> +	q =3D bio->bi_disk->queue;=0A=
>>>> +	if (!blk_queue_limit_bio_size(q))=0A=
>>>> +		return UINT_MAX;=0A=
>>>> +=0A=
>>>> +	return blk_queue_get_max_sectors(q, bio_op(bio)) << SECTOR_SHIFT;=0A=
>>>> +}=0A=
>>>> +EXPORT_SYMBOL(bio_max_size);=0A=
>>>> +=0A=
>>>>  /**=0A=
>>>>   * bio_reset - reinitialize a bio=0A=
>>>>   * @bio:	bio to reset=0A=
>>>> @@ -877,7 +892,7 @@ bool __bio_try_merge_page(struct bio *bio, struct =
page *page,=0A=
>>>>  		struct bio_vec *bv =3D &bio->bi_io_vec[bio->bi_vcnt - 1];=0A=
>>>>  =0A=
>>>>  		if (page_is_mergeable(bv, page, len, off, same_page)) {=0A=
>>>> -			if (bio->bi_iter.bi_size > UINT_MAX - len) {=0A=
>>>> +			if (bio->bi_iter.bi_size > bio_max_size(bio) - len) {=0A=
>>>>  				*same_page =3D false;=0A=
>>>>  				return false;=0A=
>>>>  			}=0A=
>>>=0A=
>>> So far we don't need bio->bi_disk or bio->bi_bdev(will be changed in=0A=
>>> Christoph's patch) during adding page to bio, so there is null ptr=0A=
>>> refereance risk.=0A=
>>>=0A=
>>>> diff --git a/include/linux/bio.h b/include/linux/bio.h=0A=
>>>> index 1edda614f7ce..cdb134ca7bf5 100644=0A=
>>>> --- a/include/linux/bio.h=0A=
>>>> +++ b/include/linux/bio.h=0A=
>>>> @@ -100,6 +100,8 @@ static inline void *bio_data(struct bio *bio)=0A=
>>>>  	return NULL;=0A=
>>>>  }=0A=
>>>>  =0A=
>>>> +extern unsigned int bio_max_size(struct bio *);=0A=
>>>> +=0A=
>>>>  /**=0A=
>>>>   * bio_full - check if the bio is full=0A=
>>>>   * @bio:	bio to check=0A=
>>>> @@ -113,7 +115,7 @@ static inline bool bio_full(struct bio *bio, unsig=
ned len)=0A=
>>>>  	if (bio->bi_vcnt >=3D bio->bi_max_vecs)=0A=
>>>>  		return true;=0A=
>>>>  =0A=
>>>> -	if (bio->bi_iter.bi_size > UINT_MAX - len)=0A=
>>>> +	if (bio->bi_iter.bi_size > bio_max_size(bio) - len)=0A=
>>>>  		return true;=0A=
>>>>  =0A=
>>>>  	return false;=0A=
>>>> diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h=0A=
>>>> index f94ee3089e01..3aeab9e7e97b 100644=0A=
>>>> --- a/include/linux/blkdev.h=0A=
>>>> +++ b/include/linux/blkdev.h=0A=
>>>> @@ -621,6 +621,7 @@ struct request_queue {=0A=
>>>>  #define QUEUE_FLAG_RQ_ALLOC_TIME 27	/* record rq->alloc_time_ns */=0A=
>>>>  #define QUEUE_FLAG_HCTX_ACTIVE	28	/* at least one blk-mq hctx is acti=
ve */=0A=
>>>>  #define QUEUE_FLAG_NOWAIT       29	/* device supports NOWAIT */=0A=
>>>> +#define QUEUE_FLAG_LIMIT_BIO_SIZE 30	/* limit bio size */=0A=
>>>>  =0A=
>>>>  #define QUEUE_FLAG_MQ_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\=0A=
>>>>  				 (1 << QUEUE_FLAG_SAME_COMP) |		\=0A=
>>>> @@ -667,6 +668,8 @@ bool blk_queue_flag_test_and_set(unsigned int flag=
, struct request_queue *q);=0A=
>>>>  #define blk_queue_fua(q)	test_bit(QUEUE_FLAG_FUA, &(q)->queue_flags)=
=0A=
>>>>  #define blk_queue_registered(q)	test_bit(QUEUE_FLAG_REGISTERED, &(q)-=
>queue_flags)=0A=
>>>>  #define blk_queue_nowait(q)	test_bit(QUEUE_FLAG_NOWAIT, &(q)->queue_f=
lags)=0A=
>>>> +#define blk_queue_limit_bio_size(q)	\=0A=
>>>> +	test_bit(QUEUE_FLAG_LIMIT_BIO_SIZE, &(q)->queue_flags)=0A=
>>>=0A=
>>> I don't think it is a good idea by adding queue flag for this purpose,=
=0A=
>>> since this case just needs to limit bio size for not delay bio submissi=
on=0A=
>>> too much, which is kind of logical thing, nothing to do with request qu=
eue.=0A=
>>>=0A=
>>> Just wondering why you don't take the following way:=0A=
>>>=0A=
>>>=0A=
>>> diff --git a/block/bio.c b/block/bio.c=0A=
>>> index 99040a7e6656..35852f7f47d4 100644=0A=
>>> --- a/block/bio.c=0A=
>>> +++ b/block/bio.c=0A=
>>> @@ -1081,7 +1081,7 @@ static int __bio_iov_append_get_pages(struct bio =
*bio, struct iov_iter *iter)=0A=
>>>   * It's intended for direct IO, so doesn't do PSI tracking, the caller=
 is=0A=
>>>   * responsible for setting BIO_WORKINGSET if necessary.=0A=
>>>   */=0A=
>>> -int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)=0A=
>>> +int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter, boo=
l sync)=0A=
>>>  {=0A=
>>>  	int ret =3D 0;=0A=
>>>  =0A=
>>> @@ -1092,12 +1092,20 @@ int bio_iov_iter_get_pages(struct bio *bio, str=
uct iov_iter *iter)=0A=
>>>  		bio_set_flag(bio, BIO_NO_PAGE_REF);=0A=
>>>  		return 0;=0A=
>>>  	} else {=0A=
>>> +		/*=0A=
>>> +		 * Don't add too many pages in case of sync dio for=0A=
>>> +		 * avoiding delay bio submission too much especially=0A=
>>> +		 * pinning user pages in memory isn't cheap.=0A=
>>> +		 */=0A=
>>> +		const unsigned int max_size =3D sync ? (1U << 12) : UINT_MAX;=0A=
>>=0A=
>> 4KB max bio size ? That is a little small :)=0A=
> =0A=
> It should have been (1U << 20), :-(=0A=
=0A=
Sounds better !=0A=
=0A=
> =0A=
>> In any case, I am not a fan of using an arbitrary value not related to t=
he=0A=
>> actual device characteristics. Wouldn't it be better to us the device=0A=
>> max_sectors limit ? And that limit would need to be zone_append_max_sect=
ors for=0A=
>> zone append writes. So some helper like Changheun bio_max_size() may be =
useful.=0A=
> =0A=
> Firstly, bio->bi_disk may not be initialized when adding page to bio; sec=
ondly this=0A=
> limit isn't really related with device, is it? Also if it is one queue li=
mit, it has=0A=
> to be stacked.=0A=
=0A=
1MB can be used as a fallback default if the gendisk is not yet set. If it =
is,=0A=
using a queue limit that does not cause bio splitting after submit makes mo=
st=0A=
sense as that avoid useless overhead. I agree, it is not critical, but it w=
ould=0A=
be nice to have some number that causes less splitting than the arbitrary 1=
MB.=0A=
E,g, most HDDs will likely have that 1MB BIO split... And max_sectors is a=
=0A=
stacked queue limit, no ? We could use max_hw_sectors too I think.=0A=
=0A=
=0A=
=0A=
-- =0A=
Damien Le Moal=0A=
Western Digital Research=0A=