Re: [PATCH v5 1/7] scsi: ufs: Add "wb_on" sysfs node to control WB on/off

From: Can Guo <cang@codeaurora.org>
To: Bean Huo <huobean@gmail.com>
Cc: Stanley Chu <stanley.chu@mediatek.com>,
	alim.akhtar@samsung.com, avri.altman@wdc.com,
	asutoshd@codeaurora.org, jejb@linux.ibm.com,
	martin.petersen@oracle.com, beanhuo@micron.com,
	bvanassche@acm.org, tomas.winkler@intel.com,
	linux-scsi@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH v5 1/7] scsi: ufs: Add "wb_on" sysfs node to control WB on/off
Date: Wed, 23 Dec 2020 09:31:49 +0800	[thread overview]
Message-ID: <862483add1462510b809aee6d3678435@codeaurora.org> (raw)
In-Reply-To: <28211d08700d1e4876a9aea342e8fcb79534cd2c.camel@gmail.com>

On 2020-12-23 06:11, Bean Huo wrote:
> On Tue, 2020-12-22 at 21:57 +0100, Bean Huo wrote:
>> > > May this operation race with UFS shutdown flow?
>> > >
>> > > To be more clear, ufshcd_wb_ctrl() here may be executed after
>> > > host
>> > > clock
>> > > is disabled by shutdown flow?
>> > >
>> > > If yes, we need to avoid it.
>> >
>> > I have the same doubt - can user still access sysfs nodes after
>> > system
>> > starts to run shutdown routines? If yes, then we need to remove all
>> > UFS
>> > sysfs nodes in ufshcd_shutdown().
>> >
>> 
>> No, we shouldn't do in this way, user space complains this. I think
>> the nodes in the sysfs can be shileded write, but the nodes shouldn't
>> be flash of its presence frequently.
>> 
>> Thanks,
>> Bean
>> 
>> 
>> > Thanks,
>> >
>> > Can Guo.
> 
> 
> Hi Can
> Got your point, you don't want user space to interrupt UFS by sysyfs if
> UFS is in power down mode. if this is true, insteading of removing all
> sysfs node in ufshcd_shutdown, maybe just add this checkup before
> accessing UFS device descriptors/flag/attributes/LU:
> 
> for example, for the device descriptor:
> 
> 
> diff --git a/drivers/scsi/ufs/ufs-sysfs.c b/drivers/scsi/ufs/ufs-
> sysfs.c
> index b3bf7fca00e5..881fe1c24a9f 100644
> --- a/drivers/scsi/ufs/ufs-sysfs.c
> +++ b/drivers/scsi/ufs/ufs-sysfs.c
> @@ -262,6 +262,9 @@ static ssize_t ufs_sysfs_read_desc_param(struct
> ufs_hba *hba,
>         u8 desc_buf[8] = {0};
>         int ret;
> 
> +       if (!ufshcd_is_ufs_dev_active(hba) ||
> !ufshcd_is_link_active(hba))
> +               return -EACCES;
> +
> 
> Bean

First of all, this check is not helping at all, during 
ufshcd_shutdown(),
both the link and dev can be active for a moment, so this check is not
helping the race condition. And assume you come up with a better check,
you want to add the check everywhere? You must have noticed the fix to
the func ufshcd_clk_gate_enable_store() from Jaegeuk Kim. So, don't
expect any luck from user space, so long the sysfs nodes are available,
users can grasp them and even stress them just for testing purposes.
I don't see why removing the sysfs nodes during ufshcd_shutdown() is a
concern to customer - we should do whatever is needed to protect LLD
contexts from user space intervene. What do you think?

Can Guo.

Thanks,

Can Guo.