Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754262AbbFQDEw (ORCPT ); Tue, 16 Jun 2015 23:04:52 -0400 Received: from smtpbg303.qq.com ([184.105.206.26]:41904 "EHLO smtpbg303.qq.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750863AbbFQDEm (ORCPT ); Tue, 16 Jun 2015 23:04:42 -0400 X-QQ-mid: bizesmtp8t1434510241t455t277 X-QQ-SSF: 0140000000200010F322B00A0000000 X-QQ-FEAT: JibMalLukFbm8Aq4CR5bZf2dVOsYh05rLPXLxKD6U0raDtGsEDjss1hO/u6D/ f+vRfQ2lluaMTLHBn8uGGp609NzRhESFnpNgsszUN5KpFzCir+nlMsB9ojrI8GYPs7/W9wi M9OYUpvtaq6uZx6ls5yp7ajPHSLaf4Rh5Ql5mIkkq0e82zjxUq1K4renNjwkEuQ8GnWmFmI 3OqVw4Rv3fj8J1IJNljUZ/5L6fLvKdCcL8jkMIjSwtw== X-QQ-GoodBg: 2 Message-ID: <5580E3A0.6050504@unitedstack.com> Date: Wed, 17 Jun 2015 11:04:00 +0800 From: juncheng bai User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:31.0) Gecko/20100101 Thunderbird/31.7.0 MIME-Version: 1.0 To: Ilya Dryomov CC: idryomov@redhat.com, Alex Elder , Josh Durgin , Guangliang Zhao , jeff@garzik.org, yehuda@hq.newdream.net, Sage Weil , elder@inktank.com, "linux-kernel@vger.kernel.org" , Ceph Development Subject: Re: [PATCH RFC] storage:rbd: make the size of request is equal to the, size of the object References: <557EB47F.6090708@unitedstack.com> <557ED1D4.20605@unitedstack.com> <557F97CB.6070608@unitedstack.com> <55800F24.6060100@unitedstack.com> <55802F46.7020804@unitedstack.com> In-Reply-To: Content-Type: text/plain; charset=utf-8; format=flowed Content-Transfer-Encoding: 7bit X-QQ-SENDSIZE: 520 X-QQ-FName: 4F1A7B99974A446587CB748B1BE51A9A X-QQ-LocalIP: 112.95.241.173 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 12400 Lines: 364 On 2015/6/16 23:51, Ilya Dryomov wrote: > On Tue, Jun 16, 2015 at 5:14 PM, juncheng bai > wrote: >> >> >> On 2015/6/16 21:30, Ilya Dryomov wrote: >>> >>> On Tue, Jun 16, 2015 at 2:57 PM, juncheng bai >>> wrote: >>>> >>>> >>>> >>>> On 2015/6/16 16:37, Ilya Dryomov wrote: >>>>> >>>>> >>>>> On Tue, Jun 16, 2015 at 6:28 AM, juncheng bai >>>>> wrote: >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> On 2015/6/15 22:27, Ilya Dryomov wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Mon, Jun 15, 2015 at 4:23 PM, juncheng bai >>>>>>> wrote: >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On 2015/6/15 21:03, Ilya Dryomov wrote: >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> On Mon, Jun 15, 2015 at 2:18 PM, juncheng bai >>>>>>>>> wrote: >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> From 6213215bd19926d1063d4e01a248107dab8a899b Mon Sep 17 >>>>>>>>>> 00:00:00 >>>>>>>>>> 2001 >>>>>>>>>> From: juncheng bai >>>>>>>>>> Date: Mon, 15 Jun 2015 18:34:00 +0800 >>>>>>>>>> Subject: [PATCH] storage:rbd: make the size of request is equal to >>>>>>>>>> the >>>>>>>>>> size of the object >>>>>>>>>> >>>>>>>>>> ensures that the merged size of request can achieve the size of >>>>>>>>>> the object. >>>>>>>>>> when merge a bio to request or merge a request to request, the >>>>>>>>>> sum of the segment number of the current request and the segment >>>>>>>>>> number of the bio is not greater than the max segments of the >>>>>>>>>> request, >>>>>>>>>> so the max size of request is 512k if the max segments of request >>>>>>>>>> is >>>>>>>>>> BLK_MAX_SEGMENTS. >>>>>>>>>> >>>>>>>>>> Signed-off-by: juncheng bai >>>>>>>>>> --- >>>>>>>>>> drivers/block/rbd.c | 2 ++ >>>>>>>>>> 1 file changed, 2 insertions(+) >>>>>>>>>> >>>>>>>>>> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c >>>>>>>>>> index 0a54c58..dec6045 100644 >>>>>>>>>> --- a/drivers/block/rbd.c >>>>>>>>>> +++ b/drivers/block/rbd.c >>>>>>>>>> @@ -3757,6 +3757,8 @@ static int rbd_init_disk(struct rbd_device >>>>>>>>>> *rbd_dev) >>>>>>>>>> segment_size = rbd_obj_bytes(&rbd_dev->header); >>>>>>>>>> blk_queue_max_hw_sectors(q, segment_size / >>>>>>>>>> SECTOR_SIZE); >>>>>>>>>> blk_queue_max_segment_size(q, segment_size); >>>>>>>>>> + if (segment_size > BLK_MAX_SEGMENTS * PAGE_SIZE) >>>>>>>>>> + blk_queue_max_segments(q, segment_size / >>>>>>>>>> PAGE_SIZE); >>>>>>>>>> blk_queue_io_min(q, segment_size); >>>>>>>>>> blk_queue_io_opt(q, segment_size); >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> I made a similar patch on Friday, investigating blk-mq plugging >>>>>>>>> issue >>>>>>>>> reported by Nick. My patch sets it to BIO_MAX_PAGES unconditionally >>>>>>>>> - >>>>>>>>> AFAIU there is no point in setting to anything bigger since the bios >>>>>>>>> will be clipped to that number of vecs. Given that BIO_MAX_PAGES is >>>>>>>>> 256, this gives is 1M direct I/Os. >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> Hi. For signal bio, the max number of bio_vec is BIO_MAX_PAGES, but a >>>>>>>> request can be merged from multiple bios. We can see the below >>>>>>>> function: >>>>>>>> ll_back_merge_fn, ll_front_merge_fn and etc. >>>>>>>> And I test in kernel 3.18 use this patch, and do: >>>>>>>> echo 4096 > /sys/block/rbd0/queue/max_sectors_kb >>>>>>>> We use systemtap to trace the request size, It is upto 4M. >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> Kernel 3.18 is pre rbd blk-mq transition, which happened in 4.0. You >>>>>>> should test whatever patches you have with at least 4.0. >>>>>>> >>>>>>> Putting that aside, I must be missing something. You'll get 4M >>>>>>> requests on 3.18 both with your patch and without it, the only >>>>>>> difference would be the size of bios being merged - 512k vs 1M. Can >>>>>>> you describe your test workload and provide before and after traces? >>>>>>> >>>>>> Hi. I update kernel version to 4.0.5. The test information as shown >>>>>> below: >>>>>> The base information: >>>>>> 03:28:13-root@server-186:~$uname -r >>>>>> 4.0.5 >>>>>> >>>>>> My simple systemtap script: >>>>>> probe module("rbd").function("rbd_img_request_create") >>>>>> { >>>>>> printf("offset:%lu length:%lu\n", ulong_arg(2), ulong_arg(3)); >>>>>> } >>>>>> >>>>>> I use dd to execute the test case: >>>>>> dd if=/dev/zero of=/dev/rbd0 bs=4M count=1 oflag=direct >>>>>> >>>>>> Case one: Without patch >>>>>> 03:30:23-root@server-186:~$cat /sys/block/rbd0/queue/max_sectors_kb >>>>>> 4096 >>>>>> 03:30:35-root@server-186:~$cat /sys/block/rbd0/queue/max_segments >>>>>> 128 >>>>>> >>>>>> The output of systemtap for nornal data: >>>>>> offset:0 length:524288 >>>>>> offset:524288 length:524288 >>>>>> offset:1048576 length:524288 >>>>>> offset:1572864 length:524288 >>>>>> offset:2097152 length:524288 >>>>>> offset:2621440 length:524288 >>>>>> offset:3145728 length:524288 >>>>>> offset:3670016 length:524288 >>>>>> >>>>>> Case two:With patch >>>>>> cat /sys/block/rbd0/queue/max_sectors_kb >>>>>> 4096 >>>>>> 03:49:14-root@server-186:linux-4.0.5$cat >>>>>> /sys/block/rbd0/queue/max_segments >>>>>> 1024 >>>>>> The output of systemtap for nornal data: >>>>>> offset:0 length:1048576 >>>>>> offset:1048576 length:1048576 >>>>>> offset:2097152 length:1048576 >>>>>> offset:3145728 length:1048576 >>>>>> >>>>>> According to the test, you are right. >>>>>> Because the blk-mq doesn't use any scheduling policy. >>>>>> 03:52:13-root@server-186:linux-4.0.5$cat >>>>>> /sys/block/rbd0/queue/scheduler >>>>>> none >>>>>> >>>>>> In previous versions of the kernel 4.0, the rbd use the defualt >>>>>> scheduler:cfq >>>>>> >>>>>> So, I think that the blk-mq need to do more? >>>>> >>>>> >>>>> >>>>> There is no scheduler support in blk-mq as of now but your numbers >>>>> don't have anything to do with that. The current behaviour is a result >>>>> of a bug in blk-mq. It's fixed by [1], if you apply it you should see >>>>> 4M requests with your stap script. >>>>> >>>>> [1] http://article.gmane.org/gmane.linux.kernel/1941750 >>>>> >>>> Hi. >>>> First, Let's look at the result in the kernel version 3.18 >>>> The function blk_limits_max_hw_sectors different implemention between >>>> 3.18 >>>> and 4.0+. We need do: >>>> echo 4094 >/sys/block/rbd0/queue/max_sectors_kb >>>> >>>> The rbd device information: >>>> 11:13:18-root@server-186:~$cat /sys/block/rbd0/queue/max_sectors_kb >>>> 4094 >>>> 11:15:28-root@server-186:~$cat /sys/block/rbd0/queue/max_segments >>>> 1024 >>>> >>>> The test command: >>>> dd if=/dev/zero of=/dev/rbd0 bs=4M count=1 >>>> >>>> The simple stap script: >>>> probe module("rbd").function("rbd_img_request_create") >>>> { >>>> printf("offset:%lu length:%lu\n", ulong_arg(2), ulong_arg(3)); >>>> } >>>> >>>> The output from stap: >>>> offset:0 length:4190208 >>>> offset:21474770944 length:4096 >>>> >>>> Second, thanks for your patch [1]. >>>> I use the patch [1], and recompile the kernel. >>>> The test information as shown below: >>>> 12:26:12-root@server-186:$cat /sys/block/rbd0/queue/max_segments >>>> 1024 >>>> 12:26:23-root@server-186:$cat /sys/block/rbd0/queue/max_sectors_kb >>>> 4096 >>>> >>>> The test command: >>>> dd if=/dev/zero of=/dev/rbd0 bs=4M count=2 oflag=direct >>>> >>>> The simple systemtap script: >>>> probe module("rbd").function("rbd_img_request_create") >>>> { >>>> printf("offset:%lu length:%lu\n", ulong_arg(2), ulong_arg(3)); >>>> } >>>> >>>> The output of systemtap for nornal data: >>>> offset:0 length:4194304 >>>> offset:4194304 length:4194304 >>>> offset:21474770944 length:4096 >>> >>> >>> Sorry, I fail to see the purpose of the above tests. The test commands >>> differ, the kernels differ and it looks like you had your patch applied >>> for both tests. What I'm trying to get you to do is to show me some >>> data that will back your claim (which your patch is based on): >>> >>>> >>>> So, I think that the max_segments of request_limits should be divide the >>>> object size by PAGE_SIZE. >>> >>> >>> For that you need to use the same kernel and run the same workload. >>> The only difference should be whether your patch is applied or not. >>> I still think that setting rbd max_segments to anything above >>> BIO_MAX_PAGES is bogus, but I'd be happy to be shown wrong on that >>> since that would mean better performance, at least in some >>> workloads. >>> >> Hi. >> For cloned image, it will avoid doing copyup if the request size is >> equal to the object size, I think that it is the key effect of this >> patch. >> The big request would result in overtime if the ceph backend is busy >> or the network bandwidth is too low. > > You are right, but then again: we get rbd object size sized requests > even with the default max_segments. This is true for both < 4.0 and >> = 4.0 kernels (with the plugging fix applied). Hi. Yeah, you are right, use the default max_segments, the request size can be the object size, because the bi_phys_segments of bio could be recount, there's just a possibility. I want to fully understand the bi_phys_segments, hope you can give me some information, thanks. The test information as shown below: The systemtap script: global greq=0; probe kernel.function("bio_attempt_back_merge") { greq=pointer_arg(2); } probe kernel.function("bio_attempt_back_merge").return { printf("after req addr:%p req segments:%d req offset:%lu req length:%lu\n", greq, @cast(greq, "request")->nr_phys_segments, @cast(greq, "request")->__sector * 512, @cast(greq, "request")->__data_len); } probe kernel.function("blk_mq_start_request") { printf("req addr:%p nr_phys_segments:%d, offset:%lu len:%lu\n", pointer_arg(1), @cast(pointer_arg(1), "request")->nr_phys_segments, @cast(pointer_arg(1), "request")->__sector * 512, @cast(pointer_arg(1), "request")->__data_len); } Test command: dd if=/dev/zero of=/dev/rbd0 bs=4M count=2 oflag=direct seek=100 Cast one: blk_queue_max_segments(q, 256); The output of stap: after req addr:0xffff880ff60a08c0 req segments:73 req offset:419430400 req length:2097152 after req addr:0xffff880ff60a08c0 req segments:73 req offset:419430400 req length:2097152 after req addr:0xffff880ff60a0a80 req segments:186 req offset:421527552 req length:1048576 req addr:0xffff880ff60a08c0 nr_phys_segments:73, offset:419430400 len:2097152 req addr:0xffff880ff60a0a80 nr_phys_segments:186, offset:421527552 len:1048576 req addr:0xffff880ff60a0c40 nr_phys_segments:232, offset:422576128 len:1048576 after req addr:0xffff880ff60a0c40 req segments:73 req offset:423624704 req length:2097152 after req addr:0xffff880ff60a0c40 req segments:73 req offset:423624704 req length:2097152 after req addr:0xffff880ff60a0e00 req segments:186 req offset:425721856 req length:1048576 req addr:0xffff880ff60a0c40 nr_phys_segments:73, offset:423624704 len:2097152 req addr:0xffff880ff60a0e00 nr_phys_segments:186, offset:425721856 len:1048576 req addr:0xffff880ff60a0fc0 nr_phys_segments:232, offset:426770432 len:1048576 Case two: blk_queue_max_segments(q, segment_size / PAGE_SIZE); The output of stap: after req addr:0xffff88101c9a0000 req segments:478 req offset:419430400 req length:4194304 req addr:0xffff88101c9a0000 nr_phys_segments:478, offset:419430400 len:4194304 after req addr:0xffff88101c9a0000 req segments:478 req offset:423624704 req length:4194304 req addr:0xffff88101c9a0000 nr_phys_segments:478, offset:423624704 len:4194304 1.Based on the setting of max_sectors and max_segments, decides the size of a request. 2.We have already set max_sectors to an object's size, so we should try to ensure that a request to the size as possible as merge bio. Thanks. ---- juncheng bai > >> I suggest that add a module parameter to control the value which >> decided by the user settings. > > A module parameter for what exactly? It likes the single_major of module rbd. > > Thanks, > > Ilya > -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/