From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-0.8 required=3.0 tests=HEADER_FROM_DIFFERENT_DOMAINS, MAILING_LIST_MULTI,SPF_PASS,URIBL_BLOCKED autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 5F307C677FC for ; Thu, 11 Oct 2018 15:39:07 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 25740213A2 for ; Thu, 11 Oct 2018 15:39:07 +0000 (UTC) DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 25740213A2 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=linux.intel.com Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=linux-kernel-owner@vger.kernel.org Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1729813AbeJKXGr (ORCPT ); Thu, 11 Oct 2018 19:06:47 -0400 Received: from mga02.intel.com ([134.134.136.20]:33590 "EHLO mga02.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1726721AbeJKXGr (ORCPT ); Thu, 11 Oct 2018 19:06:47 -0400 X-Amp-Result: SKIPPED(no attachment in message) X-Amp-File-Uploaded: False Received: from orsmga003.jf.intel.com ([10.7.209.27]) by orsmga101.jf.intel.com with ESMTP/TLS/DHE-RSA-AES256-GCM-SHA384; 11 Oct 2018 08:39:04 -0700 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.54,368,1534834800"; d="scan'208";a="91159768" Received: from ahduyck-mobl.amr.corp.intel.com (HELO [10.7.198.157]) ([10.7.198.157]) by orsmga003.jf.intel.com with ESMTP; 11 Oct 2018 08:38:56 -0700 Subject: Re: [PATCH v5 4/4] mm: Defer ZONE_DEVICE page initialization to the point where we init pgmap To: Dan Williams , Michal Hocko , Linux MM , Andrew Morton , Linux Kernel Mailing List , linux-nvdimm , Pasha Tatashin , Dave Hansen , =?UTF-8?B?SsOpcsO0bWUgR2xpc3Nl?= , rppt@linux.vnet.ibm.com, Ingo Molnar , "Kirill A. Shutemov" References: <20180925200551.3576.18755.stgit@localhost.localdomain> <20180925202053.3576.66039.stgit@localhost.localdomain> <20181009170051.GA40606@tiger-server> <25092df0-b7b4-d456-8409-9c004cb6e422@linux.intel.com> <20181010095838.GG5873@dhcp22.suse.cz> <20181010172451.GK5873@dhcp22.suse.cz> <20181011083953.GB51021@tiger-server> From: Alexander Duyck Message-ID: Date: Thu, 11 Oct 2018 08:38:56 -0700 User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:60.0) Gecko/20100101 Thunderbird/60.2.1 MIME-Version: 1.0 In-Reply-To: <20181011083953.GB51021@tiger-server> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-US Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 10/11/2018 1:39 AM, Yi Zhang wrote: > On 2018-10-10 at 11:18:49 -0700, Dan Williams wrote: >> On Wed, Oct 10, 2018 at 10:30 AM Michal Hocko wrote: >>> >>> On Wed 10-10-18 09:39:08, Alexander Duyck wrote: >>>> On 10/10/2018 2:58 AM, Michal Hocko wrote: >>>>> On Tue 09-10-18 13:26:41, Alexander Duyck wrote: >>>>> [...] >>>>>> I would think with that being the case we still probably need the call to >>>>>> __SetPageReserved to set the bit with the expectation that it will not be >>>>>> cleared for device-pages since the pages are not onlined. Removing the call >>>>>> to __SetPageReserved would probably introduce a number of regressions as >>>>>> there are multiple spots that use the reserved bit to determine if a page >>>>>> can be swapped out to disk, mapped as system memory, or migrated. >>>>> >>>>> PageReserved is meant to tell any potential pfn walkers that might get >>>>> to this struct page to back off and not touch it. Even though >>>>> ZONE_DEVICE doesn't online pages in traditional sense it makes those >>>>> pages available for further use so the page reserved bit should be >>>>> cleared. >>>> >>>> So from what I can tell that isn't necessarily the case. Specifically if the >>>> pagemap type is MEMORY_DEVICE_PRIVATE or MEMORY_DEVICE_PUBLIC both are >>>> special cases where the memory may not be accessible to the CPU or cannot be >>>> pinned in order to allow for eviction. >>> >>> Could you give me an example please? >>> >>>> The specific case that Dan and Yi are referring to is for the type >>>> MEMORY_DEVICE_FS_DAX. For that type I could probably look at not setting the >>>> reserved bit. Part of me wants to say that we should wait and clear the bit >>>> later, but that would end up just adding time back to initialization. At >>>> this point I would consider the change more of a follow-up optimization >>>> rather than a fix though since this is tailoring things specifically for DAX >>>> versus the other ZONE_DEVICE types. >>> >>> I thought I have already made it clear that these zone device hacks are >>> not acceptable to the generic hotplug code. If the current reserve bit >>> handling is not correct then give us a specific reason for that and we >>> can start thinking about the proper fix. >>> >> >> Right, so we're in a situation where a hack is needed for KVM's >> current interpretation of the Reserved flag relative to dax mapped >> pages. I'm arguing to push that knowledge / handling as deep as >> possible into the core rather than hack the leaf implementations like >> KVM, i.e. disable the Reserved flag for all non-MEMORY_DEVICE_* >> ZONE_DEVICE types. >> >> Here is the KVM thread about why they need a change: >> >> https://lkml.org/lkml/2018/9/7/552 >> >> ...and where I pushed back on a KVM-local hack: >> >> https://lkml.org/lkml/2018/9/19/154 > Yeah, Thank Dan, I think I can going on with something like this: > > @@ -5589,6 +5589,7 @@ void __ref memmap_init_zone_device(struct zone *zone, > struct page *page = pfn_to_page(pfn); > > __init_single_page(page, pfn, zone_idx, nid); > + /* Could we move this a little bit earlier as I can > + * direct use is_dax_page(page), or something else? > + */ > + page->pgmap = pgmap; > > /* > * Mark page reserved as it will need to wait for onlining > @@ -5597,14 +5598,14 @@ void __ref memmap_init_zone_device(struct zone *zone, > * We can use the non-atomic __set_bit operation for setting > * the flag as we are still initializing the pages. > */ > - __SetPageReserved(page); > + if(!is_dax_page(page)) > + __SetPageReserved(page); > > /* > * ZONE_DEVICE pages union ->lru with a ->pgmap back > * pointer and hmm_data. It is a bug if a ZONE_DEVICE > * page is ever freed or placed on a driver-private list. > */ > - page->pgmap = pgmap; > page->hmm_data = 0; > > > After Alex's patch merged. So I am not a huge fan of splitting up the pgmap init from the hmm_data, but I suppose this is just for your proof-of-concept? I already have another patch set outstanding that may actually make this change easier[1]. What I could do is add the logic there based on the pgmap.type as an additional patch since I pass a boolean to determine if I am setting the reserved bit or not. [1] https://lore.kernel.org/lkml/20181005151006.17473.83040.stgit@localhost.localdomain/