On Tue, Sep 10, 2019 at 02:03:17AM +0000, Yoshihiro Shimoda wrote:
> Hi Thierry,
> 
> > From: Thierry Reding, Sent: Tuesday, September 10, 2019 4:19 AM
> > 
> > On Mon, Sep 09, 2019 at 06:13:31PM +0200, Christoph Hellwig wrote:
> > > On Mon, Sep 09, 2019 at 02:56:56PM +0200, Thierry Reding wrote:
> > > > From: Thierry Reding <treding@nvidia.com>
> > > >
> > > > When enabling the DMA map merging capability for a queue, ensure that
> > > > the maximum segment size does not exceed the device's limit.
> > >
> > > We can't do that unfortunately.  If we use the virt_boundary setting
> > > we do aggressive merges that there is no accounting for.  So we can't
> > > limit the segment size.
> > 
> > But that's kind of the point here. My understanding is that the maximum
> > segment size in the device's DMA parameters defines the maximum size of
> > the segment that the device can handle.
> > 
> > In the particular case that I'm trying to fix, according to the SDHCI
> > specification, these devices can handle segments that are a maximum of
> > 64 KiB in size. If we allow that segment size to be exceeded, the device
> > will no longer be able to handle them.
> > 
> > > And at least for the case how devices usually do the addressing
> > > (page based on not sgl based) that needs the virt_boundary there isn't
> > > really any concept like a segment anyway.
> > 
> > I do understand that aspect of it. However, devices that do the
> > addressing this way, wouldn't they want to set the maximum segment size
> > to something large (like UINT_MAX, which many users that don't have the
> > concept of a segment already do)?
> > 
> > If you disagree, do you have any alternative proposals other than
> > reverting the offending patch? linux-next is currently broken on all
> > systems where the SDHCI controller is behind an IOMMU.
> 
> I'm sorry for causing this trouble on your environment. I have a proposal to
> resolve this issue. The mmc_host struct will have a new caps2 flag
> like MMC_CAP2_MERGE_CAPABLE and add a condition into the queue.c like below.
> 
> +	if (host->caps2 & MMC_CAP2_MERGE_CAPABLE &&
> +	    host->max_segs < MMC_DMA_MAP_MERGE_SEGMENTS &&
> 	    dma_get_merge_boundary(mmc_dev(host)))
> 		host->can_dma_map_merge = 1;
> 	else
> 		host->can_dma_map_merge = 0;
> 
> After that, all mmc controllers disable the feature as default, and if a mmc
> controller has such capable, the host driver should set the flag.
> Ulf, is such a patch acceptable for v5.4-rcN? Anyway, I'll submit such a patch later.

I'm sure that would work, but I think that's missing the point. It's not
that the setup isn't capable of merging at all. It just can't deal with
segments that are too large.

While I was debugging this, I was frequently seeing cases where the SG
was on the order of 40 entries initially and after dma_map_sg() it was
reduced to just 4 or 5.

So I think merging is still really useful if a setup supports it via an
IOMMU. I'm just not sure I see why we can't make the code respect what-
ever the maximum segment size that the driver may have configured. That
seems like it should allow us to get the best of both worlds.

Thierry