Re: [PATCH] scsi: ufs: add a quirk to disable FUA support

From: Jaegeuk Kim <jaegeuk@kernel.org>
To: Adrian Hunter <adrian.hunter@intel.com>
Cc: linux-kernel@vger.kernel.org, linux-scsi@vger.kernel.org,
	"Martin K . Petersen" <martin.petersen@oracle.com>,
	Bart Van Assche <bvanassche@acm.org>,
	Asutosh Das <asutoshd@codeaurora.org>,
	Avri Altman <avri.altman@wdc.com>, Bean Huo <beanhuo@micron.com>,
	Stanley Chu <stanley.chu@mediatek.com>,
	Can Guo <cang@codeaurora.org>
Subject: Re: [PATCH] scsi: ufs: add a quirk to disable FUA support
Date: Wed, 1 Jun 2022 10:05:45 -0700	[thread overview]
Message-ID: <YpecaXfIxZBHIcfj@google.com> (raw)
In-Reply-To: <d3038c9e-c9ec-16e9-bad4-8b1de5e23ba6@intel.com>

On 06/01, Adrian Hunter wrote:
> On 31/05/22 23:10, Jaegeuk Kim wrote:
> > UFS stack shows very low performance of FUA comparing to write and cache_flush.
> > Let's add a quirk to adjust it.
> > 
> > E.g., average latency according to the chunk size of write
> > 
> > Write(us/KB)	4	64	256	1024	2048
> > FUA		873.792	754.604	995.624	1011.67	1067.99
> > CACHE_FLUSH	824.703	712.98	800.307	1019.5	1037.37
> 
> Wouldn't it depend on how much data might be in the cache?

I've got this average latency from 100 commands of write+cache_flush vs.
write(FUA). I think the cached data should be the same as this chunk
size.

> Do you have real-world use-cases where the difference is measurable?

I'm approaching this based on 1) f2fs uses FUA for checkpoint and fsync,
and 2) iomap uses FUA for O_DIRECT|O_DSYNC case [1].

[1] https://lore.kernel.org/lkml/20220527205955.3251982-1-jaegeuk@kernel.org/

> 
> > 
> > Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
> > ---
> >  drivers/scsi/ufs/ufshcd.c | 3 +++
> >  drivers/scsi/ufs/ufshcd.h | 5 +++++
> >  2 files changed, 8 insertions(+)
> > 
> > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> > index 3f9caafa91bf..811f3467879c 100644
> > --- a/drivers/scsi/ufs/ufshcd.c
> > +++ b/drivers/scsi/ufs/ufshcd.c
> > @@ -5035,6 +5035,9 @@ static int ufshcd_slave_configure(struct scsi_device *sdev)
> >  	 */
> >  	sdev->silence_suspend = 1;
> >  
> > +	if (hba->quirks & UFSHCD_QUIRK_BROKEN_FUA)
> > +		sdev->broken_fua = 1;
> > +
> >  	ufshcd_crypto_register(hba, q);
> >  
> >  	return 0;
> > diff --git a/drivers/scsi/ufs/ufshcd.h b/drivers/scsi/ufs/ufshcd.h
> > index 94f545be183a..6c480c6741d6 100644
> > --- a/drivers/scsi/ufs/ufshcd.h
> > +++ b/drivers/scsi/ufs/ufshcd.h
> > @@ -602,6 +602,11 @@ enum ufshcd_quirks {
> >  	 * support physical host configuration.
> >  	 */
> >  	UFSHCD_QUIRK_SKIP_PH_CONFIGURATION		= 1 << 16,
> > +
> > +	/*
> > +	 * This quirk disables FUA support.
> > +	 */
> > +	UFSHCD_QUIRK_BROKEN_FUA				= 1 << 17,
> 
> Wouldn't it be more appropriate to make it a UFS_DEVICE_QUIRK_
> since it presumably depends on the UFS device not the host controller?
> 
> Also, as already commented by others, there needs to be a user of
> the quirk

Since I asked SoC vendors can verify the performance with this quirk,
I need to wait for their reply. Meanwhile, I'm willing to disable FUA in Pixel
devices, which I cannot post any patch directly to LKML.

Agreed that, if there's no other user in upstream, I'm okay to drop
this.

> 
> >  };
> >  
> >  enum ufshcd_caps {