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;
Subject: Re: [PATCH] mm/vmscan: don't scan adjust too much if current is not
 kswapd
From:   Hongchen Zhang <zhanghongchen@loongson.cn>
To:     Mel Gorman <mgorman@suse.de>
Cc:     Matthew Wilcox <willy@infradead.org>,
        Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org
References: <20220914023318.549118-1-zhanghongchen@loongson.cn>
 <20220914155142.bf388515a39fb45bae987231@linux-foundation.org>
 <6bcb4883-03d0-88eb-4c42-84fff0a9a141@loongson.cn>
 <YyLUGnqtZXn4MjJF@casper.infradead.org>
 <54813a74-cc0e-e470-c632-78437a0d0ad4@loongson.cn>
 <YyLpls9/t6LKQefS@casper.infradead.org>
 <b52b3f49-ebf5-6f63-da1a-f57711c3f97d@loongson.cn>
 <YyQ2m9vU/plyBNas@casper.infradead.org>
 <4bd0012e-77ff-9d0d-e295-800471994aeb@loongson.cn>
Message-ID: <c3f4d1bb-418c-fbf5-c251-fd448f4d4e86@loongson.cn>
Date:   Tue, 20 Sep 2022 10:23:05 +0800
User-Agent: Mozilla/5.0 (X11; Linux loongarch64; rv:68.0) Gecko/20100101
 Thunderbird/68.7.0
MIME-Version: 1.0
In-Reply-To: <4bd0012e-77ff-9d0d-e295-800471994aeb@loongson.cn>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Precedence: bulk

Hi Mel,

The scan adjust algorithm was originally introduced by you from
commmit e82e0561dae9 ("mm: vmscan: obey proportional scanning 
requirements for kswapd"), any suggestion about this fix patch?
In short, proportional scanning is not friendly to processes other than 
kswapd.

Thanks
Hongchen Zhang

On 2022/9/16 pm 6:19, Hongchen Zhang wrote:
> Hi Andrew and Matthew,
> 
> On 2022/9/16 pm 4:40, Matthew Wilcox wrote:
>> On Fri, Sep 16, 2022 at 08:57:50AM +0800, Hongchen Zhang wrote:
>>> Hi Andrew ,
>>>
>>> On 2022/9/15 pm 5:00, Matthew Wilcox wrote:
>>>> On Thu, Sep 15, 2022 at 04:02:41PM +0800, Hongchen Zhang wrote:
>>>>> Hi Matthew,
>>>>> On 2022/9/15 pm 3:28, Matthew Wilcox wrote:
>>>>>> On Thu, Sep 15, 2022 at 09:19:48AM +0800, Hongchen Zhang wrote:
>>>>>>> [ 3748.453561] INFO: task float_bessel:77920 blocked for more 
>>>>>>> than 120
>>>>>>> seconds.
>>>>>>> [ 3748.460839]       Not tainted 5.15.0-46-generic #49-Ubuntu
>>>>>>> [ 3748.466490] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" 
>>>>>>> disables
>>>>>>> this message.
>>>>>>> [ 3748.474618] task:float_bessel    state:D stack:    0 pid:77920 
>>>>>>> ppid:
>>>>>>> 77327 flags:0x00004002
>>>>>>> [ 3748.483358] Call Trace:
>>>>>>> [ 3748.485964]  <TASK>
>>>>>>> [ 3748.488150]  __schedule+0x23d/0x590
>>>>>>> [ 3748.491804]  schedule+0x4e/0xc0
>>>>>>> [ 3748.495038]  rwsem_down_read_slowpath+0x336/0x390
>>>>>>> [ 3748.499886]  ? copy_user_enhanced_fast_string+0xe/0x40
>>>>>>> [ 3748.505181]  down_read+0x43/0xa0
>>>>>>> [ 3748.508518]  do_user_addr_fault+0x41c/0x670
>>>>>>> [ 3748.512799]  exc_page_fault+0x77/0x170
>>>>>>> [ 3748.516673]  asm_exc_page_fault+0x26/0x30
>>>>>>> [ 3748.520824] RIP: 0010:copy_user_enhanced_fast_string+0xe/0x40
>>>>>>> [ 3748.526764] Code: 89 d1 c1 e9 03 83 e2 07 f3 48 a5 89 d1 f3 a4 
>>>>>>> 31 c0 0f
>>>>>>> 01 ca c3 cc cc cc cc 0f 1f 00 0f 01 cb 83 fa 40 0f 82 70 ff ff ff 
>>>>>>> 89 d1 <f3>
>>>>>>> a4 31 c0 0f 01 ca c3 cc cc cc cc 66 08
>>>>>>> [ 3748.546120] RSP: 0018:ffffaa9248fffb90 EFLAGS: 00050206
>>>>>>> [ 3748.551495] RAX: 00007f99faa1a010 RBX: ffffaa9248fffd88 RCX:
>>>>>>> 0000000000000010
>>>>>>> [ 3748.558828] RDX: 0000000000001000 RSI: ffff9db397ab8ff0 RDI:
>>>>>>> 00007f99faa1a000
>>>>>>> [ 3748.566160] RBP: ffffaa9248fffbf0 R08: ffffcc2fc2965d80 R09:
>>>>>>> 0000000000000014
>>>>>>> [ 3748.573492] R10: 0000000000000000 R11: 0000000000000014 R12:
>>>>>>> 0000000000001000
>>>>>>> [ 3748.580858] R13: 0000000000001000 R14: 0000000000000000 R15:
>>>>>>> ffffaa9248fffd98
>>>>>>> [ 3748.588196]  ? copy_page_to_iter+0x10e/0x400
>>>>>>> [ 3748.592614]  filemap_read+0x174/0x3e0
>>>>>>
>>>>>> Interesting; it wasn't the process itself which triggered the page
>>>>>> fault; the process called read() and the kernel took the page 
>>>>>> fault to
>>>>>> satisfy the read() call.
>>>>>>
>>>>>>> [ 3748.596354]  ? ima_file_check+0x6a/0xa0
>>>>>>> [ 3748.600301]  generic_file_read_iter+0xe5/0x150
>>>>>>> [ 3748.604884]  ext4_file_read_iter+0x5b/0x190
>>>>>>> [ 3748.609164]  ? aa_file_perm+0x102/0x250
>>>>>>> [ 3748.613125]  new_sync_read+0x10d/0x1a0
>>>>>>> [ 3748.617009]  vfs_read+0x103/0x1a0
>>>>>>> [ 3748.620423]  ksys_read+0x67/0xf0
>>>>>>> [ 3748.623743]  __x64_sys_read+0x19/0x20
>>>>>>> [ 3748.627511]  do_syscall_64+0x59/0xc0
>>>>>>> [ 3748.631203]  ? syscall_exit_to_user_mode+0x27/0x50
>>>>>>> [ 3748.636144]  ? do_syscall_64+0x69/0xc0
>>>>>>> [ 3748.639992]  ? exit_to_user_mode_prepare+0x96/0xb0
>>>>>>> [ 3748.644931]  ? irqentry_exit_to_user_mode+0x9/0x20
>>>>>>> [ 3748.649872]  ? irqentry_exit+0x1d/0x30
>>>>>>> [ 3748.653737]  ? exc_page_fault+0x89/0x170
>>>>>>> [ 3748.657795]  entry_SYSCALL_64_after_hwframe+0x61/0xcb
>>>>>>> [ 3748.663030] RIP: 0033:0x7f9a852989cc
>>>>>>> [ 3748.666713] RSP: 002b:00007f9a8497dc90 EFLAGS: 00000246 ORIG_RAX:
>>>>>>> 0000000000000000
>>>>>>> [ 3748.674487] RAX: ffffffffffffffda RBX: 00007f9a8497f5c0 RCX:
>>>>>>> 00007f9a852989cc
>>>>>>> [ 3748.681817] RDX: 0000000000027100 RSI: 00007f99faa18010 RDI:
>>>>>>> 0000000000000061
>>>>>>> [ 3748.689150] RBP: 00007f9a8497dd60 R08: 0000000000000000 R09:
>>>>>>> 00007f99faa18010
>>>>>>> [ 3748.696493] R10: 0000000000000000 R11: 0000000000000246 R12:
>>>>>>> 00007f99faa18010
>>>>>>> [ 3748.703841] R13: 00005605e11c406f R14: 0000000000000001 R15:
>>>>>>> 0000000000027100
>>>>>>
>>>>>> ORIG_RAX is 0, which matches sys_read.
>>>>>> RDI is file descriptor 0x61
>>>>>> RSI is plausibly a userspace pointer, 0x7f99faa18010
>>>>>> RDX is the length, 0x27100 or 160kB.
>>>>>>
>>>>>> That all seems reasonable.
>>>>>>
>>>>>> What I really want to know is who is _holding_ the lock.  We stash
>>>>>> a pointer to the task_struct in 'owner', so we could clearly find 
>>>>>> this
>>>>>> out in the 'blocked for too long' report, and print their stack 
>>>>>> trace.
>>>>>>
>>>>> As described in the comment for __rwsem_set_reader_owned,it is hard 
>>>>> to track
>>>>> read owners.So we could not clearly find out who blocked the 
>>>>> process,it was
>>>>> caused by multiple tasks.
>>>>
>>>> Readers don't block readers.  You have a reader here, so it's being
>>>> blocked by a writer.  And that writer's task_struct is stashed in
>>>> rwsem->owner.  It would be nice if we dumped that information
>>>> automatically ... but we don't do that today.  Perhaps you could
>>>> grab that information from a crash dump if you have one.
>>>>
>>>>>> You must have done something like this already in order to deduce 
>>>>>> that
>>>>>> it was the direct reclaim path that was the problem?
>>>>>>
>>>>> The method we used is to track the direct reclaim using the
>>>>> trace_mm_vmscan_direct_reclaim_{begin,end} interface.When the problem
>>>>> occurred,we could get a very large "nr_reclaimed" which is not a 
>>>>> desirable
>>>>> value for process except kswapd.
>>>>
>>>> I disagree.  If a process needs to allocate memory then it should be
>>>> paying the cost of reclaiming that memory itself.  kswapd is a last
>>>> resort to reclaim memory when we have a workload (eg a network router)
>>>> that does its memory allocation primarily in interrupt context.
>>>>
>>> What's your opinion about this scan adjust issue? Is there a better 
>>> way to
>>> fix this issue?
>>
>> Yes, but we need you to gather more information about what's causing
>> the issue before we can suggest what that is.
>>
> I think the following scenery triggers the scan adjust issue:
> In function shrink_lruvec, we call get_scan_count and get the following 
> values:
> targets[LRU_INACTIVE_ANON]=50000
> targets[LRU_ACTIVE_ANON]=50000
> targets[LRU_INACTIVE_FILE]=128
> targets[LRU_ACTIVE_FILE]=129
> 
> After the first scan, we get more than nr_to_reclaim pages, but the 
> percentage of scanning nr[LRU_INACTIVE_FILE+LRU_ACTIVE_FILE] is 256/257,
> Then when we scan adjust, we must scan(possibly reclaim) 
> 256*(50000+50000)/257-256=99354 pages, which is too large and would 
> waste too many time.
> If it is not kswapd, it is unacceptable to reclaim so many pages.
> So we should limit the number of pages of scan adjust.
> 
> Thanks
> Hongchen Zhang