From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andrey Ryabinin Subject: =?UTF-8?B?UmU6IOetlOWkjTog562U5aSNOiBbUEFUQ0hdIG1tL21lbWNvbnRyb2wu?= =?UTF-8?Q?c:_speed_up_to_force_empty_a_memory_cgroup?= Date: Wed, 21 Mar 2018 01:08:02 +0300 Message-ID: <56508bd0-e8d7-55fd-5109-c8dacf26b13e@virtuozzo.com> References: <1521448170-19482-1-git-send-email-lirongqing@baidu.com> <20180319085355.GQ23100@dhcp22.suse.cz> <2AD939572F25A448A3AE3CAEA61328C23745764B@BC-MAIL-M28.internal.baidu.com> <20180319103756.GV23100@dhcp22.suse.cz> <2AD939572F25A448A3AE3CAEA61328C2374589DC@BC-MAIL-M28.internal.baidu.com> <20180320083950.GD23100@dhcp22.suse.cz> Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Return-path: DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=virtuozzo.com; s=selector1; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version; bh=MpvNwWwagnePeXMpWrVY8sMBCtPzlsn34NnYhjp1eCk=; b=bTGe/0atQGXTWo2Cy+jyJiEcs9YXe8ODFZin/fOdqVvjNKFOX3LPOTuIkvMqhEwU3B01Hn77hTuzJrOyCgfz8FD+Awuf8uW4WuJcV6RrNyPhWv0Mm4DwLFq6NYt9VmFv620xHRjTEXWo5CSmLjX0WHp15YcY5LHD1Z05X3UsQDY= In-Reply-To: Content-Language: en-US Sender: linux-kernel-owner@vger.kernel.org List-ID: Content-Type: text/plain; charset="us-ascii" To: David Rientjes , Michal Hocko Cc: "Li,Rongqing" , "linux-kernel@vger.kernel.org" , "linux-mm@kvack.org" , "cgroups@vger.kernel.org" , "hannes@cmpxchg.org" On 03/20/2018 11:29 PM, David Rientjes wrote: > On Tue, 20 Mar 2018, Michal Hocko wrote: > >>>>>> Although SWAP_CLUSTER_MAX is used at the lower level, but the call >>>>>> stack of try_to_free_mem_cgroup_pages is too long, increase the >>>>>> nr_to_reclaim can reduce times of calling >>>>>> function[do_try_to_free_pages, shrink_zones, hrink_node ] >>>>>> >>>>>> mem_cgroup_resize_limit >>>>>> --->try_to_free_mem_cgroup_pages: .nr_to_reclaim = max(1024, >>>>>> --->SWAP_CLUSTER_MAX), >>>>>> ---> do_try_to_free_pages >>>>>> ---> shrink_zones >>>>>> --->shrink_node >>>>>> ---> shrink_node_memcg >>>>>> ---> shrink_list <-------loop will happen in this place >>>>> [times=1024/32] >>>>>> ---> shrink_page_list >>>>> >>>>> Can you actually measure this to be the culprit. Because we should rethink >>>>> our call path if it is too complicated/deep to perform well. >>>>> Adding arbitrary batch sizes doesn't sound like a good way to go to me. >>>> >>>> Ok, I will try >>>> >>> >>> Looping in mem_cgroup_resize_limit(), which takes memcg_limit_mutex on >>> every iteration which contends with lowering limits in other cgroups (on >>> our systems, thousands), calling try_to_free_mem_cgroup_pages() with less >>> than SWAP_CLUSTER_MAX is lame. >> >> Well, if the global lock is a bottleneck in your deployments then we >> can come up with something more clever. E.g. per hierarchy locking >> or even drop the lock for the reclaim altogether. If we reclaim in >> SWAP_CLUSTER_MAX then the potential over-reclaim risk quite low when >> multiple users are shrinking the same (sub)hierarchy. >> > > I don't believe this to be a bottleneck if nr_pages is increased in > mem_cgroup_resize_limit(). > >>> It would probably be best to limit the >>> nr_pages to the amount that needs to be reclaimed, though, rather than >>> over reclaiming. >> >> How do you achieve that? The charging path is not synchornized with the >> shrinking one at all. >> > > The point is to get a better guess at how many pages, up to > SWAP_CLUSTER_MAX, that need to be reclaimed instead of 1. > >>> If you wanted to be invasive, you could change page_counter_limit() to >>> return the count - limit, fix up the callers that look for -EBUSY, and >>> then use max(val, SWAP_CLUSTER_MAX) as your nr_pages. >> >> I am not sure I understand >> > > Have page_counter_limit() return the number of pages over limit, i.e. > count - limit, since it compares the two anyway. Fix up existing callers > and then clamp that value to SWAP_CLUSTER_MAX in > mem_cgroup_resize_limit(). It's a more accurate guess than either 1 or > 1024. > JFYI, it's never 1, it's always SWAP_CLUSTER_MAX. See try_to_free_mem_cgroup_pages(): .... struct scan_control sc = { .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),