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;
From:   Hou Tao <houtao@huaweicloud.com>
Subject: Re: [PATCH] fscache: Use wake_up_var() to wake up pending volume
 acquisition
To:     David Howells <dhowells@redhat.com>
Cc:     linux-cachefs@redhat.com, Jeff Layton <jlayton@kernel.org>,
        linux-erofs@lists.ozlabs.org, linux-kernel@vger.kernel.org,
        "houtao1@huawei.com" <houtao1@huawei.com>,
        Jingbo Xu <jefflexu@linux.alibaba.com>
References: <42b33792-50e9-77d7-4d3e-ac5ce1adeda6@huaweicloud.com>
 <20221128031929.3918348-1-houtao@huaweicloud.com>
 <2308529.1670585211@warthog.procyon.org.uk>
Message-ID: <e6a23517-63fa-6381-1f2b-80e764252e0b@huaweicloud.com>
Date:   Thu, 15 Dec 2022 16:28:32 +0800
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:78.0) Gecko/20100101
 Thunderbird/78.6.0
MIME-Version: 1.0
In-Reply-To: <2308529.1670585211@warthog.procyon.org.uk>
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit
Content-Language: en-US
Precedence: bulk

Hi David,

Sorry for the late reply. Busy for other business in work.

On 12/9/2022 7:26 PM, David Howells wrote:
> Hou Tao <houtao@huaweicloud.com> wrote:
>
>>>  			clear_bit(FSCACHE_VOLUME_ACQUIRE_PENDING, &cursor->flags);
> Maybe this should be clear_bit_unlock() instead.
I'm not sure about that. In my understanding, clear_bit_unlock() is usually
paired with test_and_set_bit_lock() to implement bit lock to make sure the
writes before clear_bit_unlock() are visible to read access in concurrent
process, right ? But now the caller of fscache_wake_pending_volume() only modify
cursor->flags and nothing else, so I don't think it is needed here.
If its intended purpose is to provide the missing smp_mb() for wake_up_bit(), I
also don't think it is right, because the release barrier provided by
clear_bit_unlock() doesn't guarantee the order of cursor->flags and wq_head, so
I think one extra smp_mb_after_atomic() is also needed after
clear_bit(FSCACHE_VOLUME_ACQUIRE_PENDING, &cursor->flags).

If the above reasoning makes sense to you, I think we also need to add
smp_mb_after_atomic() for wake_up_bit() in fscache_create_volume_work().
> And I wonder if:
>
> 	set_bit(FSCACHE_VOLUME_ACQUIRE_PENDING, &candidate->flags);
>
> in fscache_hash_volume() needs a barrier before it.
I also don't get it. The barrier is used to guarantee the order between
cursor->flags and candidate->flags, right ? But the write and read of
cursor->flags and candidate->flags are protected by the same hash lock.
>
>>> -			wake_up_bit(&cursor->flags, FSCACHE_VOLUME_ACQUIRE_PENDING);
>>> +			/*
>>> +			 * Paired with barrier in wait_var_event(). Check
>>> +			 * waitqueue_active() and wake_up_var() for details.
>>> +			 */
>>> +			smp_mb__after_atomic();
>>> +			wake_up_var(&cursor->flags);
> That doesn't seem right.
>
> wake_up_bit() is more selective, so should be preferred to wake_up_var().
OK. Will update fscache_wait_on_volume_collision() to use wait_on_bit() accordingly.
> David
>
>
> .