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: <13b24b09-aebd-4cfc-c45a-a08ec6ead2cf@huawei.com>
Date:   Mon, 20 Jun 2022 08:40:43 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:91.0) Gecko/20100101
 Thunderbird/91.6.1
Subject: Re: [PATCH 3/4] scsi: pm8001: Use non-atomic bitmap ops for tag alloc
 + free
To:     Damien Le Moal <damien.lemoal@opensource.wdc.com>,
        Hannes Reinecke <hare@suse.de>, <jinpu.wang@cloud.ionos.com>,
        <jejb@linux.ibm.com>, <martin.petersen@oracle.com>
CC:     <linux-scsi@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
        <Ajish.Koshy@microchip.com>
References: <1654879602-33497-1-git-send-email-john.garry@huawei.com>
 <1654879602-33497-4-git-send-email-john.garry@huawei.com>
 <303bbfad-edde-1197-679e-4a09175fb1f3@suse.de>
 <594ac0c9-a55b-bec7-77e3-a6c7e9525f6b@opensource.wdc.com>
From:   John Garry <john.garry@huawei.com>
In-Reply-To: <594ac0c9-a55b-bec7-77e3-a6c7e9525f6b@opensource.wdc.com>
Content-Type: text/plain; charset="UTF-8"; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 20/06/2022 07:07, Damien Le Moal wrote:
> On 6/20/22 15:00, Hannes Reinecke wrote:
>> On 6/10/22 18:46, John Garry wrote:
>>> In pm8001_tag_alloc() we don't require atomic set_bit() as we are already
>>> in atomic context. In pm8001_tag_free() we should use the same host
>>> spinlock to protect clearing the tag (and then don't require the atomic
>>> clear_bit()).
>>>
>>> Signed-off-by: John Garry <john.garry@huawei.com>
>>> ---
>>>    drivers/scsi/pm8001/pm8001_sas.c | 10 +++++++---
>>>    1 file changed, 7 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/scsi/pm8001/pm8001_sas.c b/drivers/scsi/pm8001/pm8001_sas.c
>>> index 3a863d776724..8e3f2f9ddaac 100644
>>> --- a/drivers/scsi/pm8001/pm8001_sas.c
>>> +++ b/drivers/scsi/pm8001/pm8001_sas.c
>>> @@ -66,7 +66,11 @@ static int pm8001_find_tag(struct sas_task *task, u32 *tag)
>>>    void pm8001_tag_free(struct pm8001_hba_info *pm8001_ha, u32 tag)
>>>    {
>>>    	void *bitmap = pm8001_ha->tags;
>>> -	clear_bit(tag, bitmap);
>>> +	unsigned long flags;
>>> +
>>> +	spin_lock_irqsave(&pm8001_ha->bitmap_lock, flags);
>>> +	__clear_bit(tag, bitmap);
>>> +	spin_unlock_irqrestore(&pm8001_ha->bitmap_lock, flags);
>>>    }
>>>    
>> This spin lock is pretty much pointless; clear_bit() is always atomic.
> 
> But __clear_bit() is not atomic. I think it was the point of this patch,
> to not use atomics and use the spinlock instead to protect bitmap.
> 
> Before the patch, pm8001_tag_alloc() takes the spinlock *and* use the
> atomic set_bit(), which is an overkill. pm8001_tag_free() only clears the
> bit using the the atomic clear_bit().

Right, so I could change to use __set_bit() in pm8001_find_tag(), but 
rather use spinlock always.

> 
> After the patch, spinlock guarantees atomicity for both alloc and free.
> 
> Not sure there is any gain from this.

A few more points to note:
- On architectures which do not support atomic operations natively, they 
have to use global spinlocks to create atomic context before doing 
non-atomic bit clearing - see atomic64.c . As such, it's better to use 
the already available pm8001_ha->bitmap_lock.
- spinlock does more than create atomic context, but also has barrier 
semantics, so proper to use consistently for protecting the same region.

Thanks,
John