Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCHv2 1/1] ext4: mballoc: Use this_cpu_read instead of
 this_cpu_ptr
To:     Hillf Danton <hdanton@sina.com>
Cc:     linux-ext4@vger.kernel.org, jack@suse.com, tytso@mit.edu,
        Markus Elfring <Markus.Elfring@web.de>,
        linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        Borislav Petkov <bp@alien8.de>,
        Marek Szyprowski <m.szyprowski@samsung.com>,
        syzbot+82f324bb69744c5f6969@syzkaller.appspotmail.com
References: <20200609123716.16888-1-hdanton@sina.com>
From:   Ritesh Harjani <riteshh@linux.ibm.com>
Date:   Wed, 10 Jun 2020 07:36:09 +0530
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101
 Thunderbird/68.8.0
MIME-Version: 1.0
In-Reply-To: <20200609123716.16888-1-hdanton@sina.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 7bit
Message-Id: <20200610020610.AEB3211C066@d06av25.portsmouth.uk.ibm.com>
Sender: linux-ext4-owner@vger.kernel.org
Precedence: bulk


On 6/9/20 6:07 PM, Hillf Danton wrote:
> 
> On Tue, 9 Jun 2020 18:53:23 +0800 Ritesh Harjani wrote:
>>
>> Simplify reading a seq variable by directly using this_cpu_read API
>> instead of doing this_cpu_ptr and then dereferencing it.
> 
> Two of the quick questions
> 1) Why can blocks discarded in a ext4 FS help allocators in another?

I am not sure if I understand your Q correctly. But here is a brief 
about the patchset. If there were PA blocks just or about to be 
discarded by another thread, then the current thread who is doing block 
allocation should not fail with ENOSPC error instead should be able to 
allocate those blocks from another thread. The concept is better 
explained in the commit msgs, if more details are required.
Without this patchset (in some heavy multi-threaded use case) allocation 
was failing when the overall filesystem space available was more then 50%.

> 
> 2) Why is a percpu seqcount prefered over what <linux/seqlock.h>
> can offer?
> 

Since this could be a multi-threaded use case, per cpu variable helps in 
avoid cache line bouncing problem, which could happen when the same 
variable is updated by multiple threads on different cpus.

-ritesh