From mboxrd@z Thu Jan  1 00:00:00 1970
From: keith.busch@intel.com (Keith Busch)
Date: Wed, 3 Jan 2018 13:11:31 -0700
Subject: Bug report - drivers/nvme/host/pci.c?
In-Reply-To: <20180103200140.GA11172@localhost.localdomain>
References: <4dcd8c76-872e-69e3-3cbd-d051bdf0bf62@huawei.com>
 <BY2PR04MB2101903E77826A0479550572861E0@BY2PR04MB2101.namprd04.prod.outlook.com>
 <20180103200140.GA11172@localhost.localdomain>
Message-ID: <20180103201131.GB11172@localhost.localdomain>

On Wed, Jan 03, 2018@01:01:41PM -0700, Keith Busch wrote:
> On Wed, Jan 03, 2018@07:46:05PM +0000, Chaitanya Kulkarni wrote:
> >    HI Fangjian,
> > 
> >    I ran your test on the latest kernel and I did not see the problem.
> <> 
> >    Is there a particular reason why you are using Linux version 4.15.0-rc1?
> > 
> >    Can you please check if you are having the same problem on the latest
> >    kernel as mentioned in my log?
> > 
> >    Regards,
> > 
> >    Chaitanya
> 
> I would guess the reporter has an IOMMU enabled that DMA maps into fewer
> entries than the virtual table has. Since that's possible, I think the
> WARN_ON is invalid unless you overwrite iod->nents to be the return of
> dma_map_sg, I think.

Just fyi, the driver selecting when to use SGL vs PRP is done prior to
DMA mapping the scatter list. An IOMMU may map a virtually contiguous
buffer into a single address, which is probably where biggest win for
SGL over PRP exists. I think this driver handling ought to be reworked.