Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <5f340d98-4ee0-35a9-58ed-943834d68042@linux.alibaba.com>
Date:   Thu, 15 Jun 2023 11:59:00 +0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101
 Thunderbird/102.11.2
Subject: Re: [PATCH v2] mm: compaction: skip memory hole rapidly when
 isolating migratable pages
To:     "Huang, Ying" <ying.huang@intel.com>,
        Mel Gorman <mgorman@techsingularity.net>, david@redhat.com
Cc:     akpm@linux-foundation.org, vbabka@suse.cz, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
References: <770f9f61472b24b6bc89adbd71a77d9cf62bb54f.1686646361.git.baolin.wang@linux.alibaba.com>
 <20230614095501.m4porztaibchrgwx@techsingularity.net>
 <87ilbpo1d9.fsf@yhuang6-desk2.ccr.corp.intel.com>
From:   Baolin Wang <baolin.wang@linux.alibaba.com>
In-Reply-To: <87ilbpo1d9.fsf@yhuang6-desk2.ccr.corp.intel.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 6/15/2023 11:22 AM, Huang, Ying wrote:
> Hi, Mel,
> 
> Mel Gorman <mgorman@techsingularity.net> writes:
> 
>> On Tue, Jun 13, 2023 at 04:55:04PM +0800, Baolin Wang wrote:
>>> On some machines, the normal zone can have a large memory hole like
>>> below memory layout, and we can see the range from 0x100000000 to
>>> 0x1800000000 is a hole. So when isolating some migratable pages, the
>>> scanner can meet the hole and it will take more time to skip the large
>>> hole. From my measurement, I can see the isolation scanner will take
>>> 80us ~ 100us to skip the large hole [0x100000000 - 0x1800000000].
>>>
>>> So adding a new helper to fast search next online memory section
>>> to skip the large hole can help to find next suitable pageblock
>>> efficiently. With this patch, I can see the large hole scanning only
>>> takes < 1us.
>>>
>>> [    0.000000] Zone ranges:
>>> [    0.000000]   DMA      [mem 0x0000000040000000-0x00000000ffffffff]
>>> [    0.000000]   DMA32    empty
>>> [    0.000000]   Normal   [mem 0x0000000100000000-0x0000001fa7ffffff]
>>> [    0.000000] Movable zone start for each node
>>> [    0.000000] Early memory node ranges
>>> [    0.000000]   node   0: [mem 0x0000000040000000-0x0000000fffffffff]
>>> [    0.000000]   node   0: [mem 0x0000001800000000-0x0000001fa3c7ffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa3c80000-0x0000001fa3ffffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa4000000-0x0000001fa402ffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa4030000-0x0000001fa40effff]
>>> [    0.000000]   node   0: [mem 0x0000001fa40f0000-0x0000001fa73cffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa73d0000-0x0000001fa745ffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa7460000-0x0000001fa746ffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa7470000-0x0000001fa758ffff]
>>> [    0.000000]   node   0: [mem 0x0000001fa7590000-0x0000001fa7ffffff]
>>>
>>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>>
>> This may only be necessary for non-contiguous zones so a check for
>> zone_contiguous could be made but I suspect the saving, if any, would be
>> marginal.
>>
>> However, it's subtle that block_end_pfn can end up in an arbirary location
>> past the end of the zone or past cc->free_pfn. As the "continue" will update
>> cc->migrate_pfn, that might lead to errors in the future. It would be a
>> lot safer to pass in cc->free_pfn and do two things with the value. First,
>> there is no point scanning for a valid online section past cc->free_pfn so
>> terminating after cc->free_pfn may save some cycles. Second, cc->migrate_pfn
>> does not end up with an arbitrary value which is a more defensive approach
>> to any future programming errors.
> 
> I have thought about this before.  Originally, I had thought that we
> were safe because cc->free_pfn should be in a online section and
> block_end_pfn should reach cc->free_pfn before the end of zone.  But
> after checking more code and thinking about it again, I found that the
> underlying sections may go offline under us during compaction.  So that,
> cc->free_pfn may be in a offline section or after the end of zone.  So,
> you are right, we need to consider the range of block_end_pfn.
> 
> But, if we thought in this way (memory online/offline at any time), it
> appears that we need to check whether the underlying section was
> offlined.  For example, is it safe to use "pfn_to_page()" in
> "isolate_migratepages_block()"?  Is it possible for the underlying
> section to be offlined under us?

It is possible. There is a previous discussion[1] about the race between 
pfn_to_online_page() and memory offline.

[1] 
https://lore.kernel.org/lkml/87zgc6buoq.fsf@nvidia.com/T/#m642d91bcc726437e1848b295bc57ce249c7ca399