RE: [PATCH v2] iommu/amd: Disable IOMMU on Stoney Ridge systems

[Deucher, Alexander]

> -----Original Message-----
> From: Lucas Stach <dev@xxxxxxxxxx>
> Sent: Sunday, December 1, 2019 7:43 AM
> To: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>; joro@xxxxxxxxxx
> Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>;
> iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH v2] iommu/amd: Disable IOMMU on Stoney Ridge
> systems
> 
> Am Freitag, den 29.11.2019, 22:21 +0800 schrieb Kai-Heng Feng:
> > Serious screen flickering when Stoney Ridge outputs to a 4K monitor.
> >
> > According to Alex Deucher, IOMMU isn't enabled on Windows, so let's do
> > the same here to avoid screen flickering on 4K monitor.
> 
> This doesn't seem like a good solution, especially if there isn't a method for
> the user to opt-out.  Some users might prefer having the IOMMU support to
> 4K display output.
> 
> But before using the big hammer of disabling or breaking one of those
> features, we should take a look at what's the issue here. Screen flickering
> caused by the IOMMU being active hints to the IOMMU not being able to
> sustain the translation bandwidth required by the high- bandwidth
> isochronous transfers caused by 4K scanout, most likely due to insufficient
> TLB space.
> 
> As far as I know the framebuffer memory for the display buffers is located in
> stolen RAM, and thus contigous in memory. I don't know the details of the
> GPU integration on those APUs, but maybe there even is a way to bypass the
> IOMMU for the stolen VRAM regions?
> 
> If there isn't and all GPU traffic passes through the IOMMU when active, we
> should check if the stolen RAM is mapped with hugepages on the IOMMU
> side. All the stolen RAM can most likely be mapped with a few hugepage
> mappings, which should reduce IOMMU TLB demand by a large margin.

The is no issue when we scan out of the carve out region.  The issue occurs when we scan out of regular system memory (scatter/gather).  Many newer laptops have very small carve out regions (e.g., 32 MB), so we have to use regular system pages to support multiple high resolution displays.  The problem is, the latency gets too high at some point when the IOMMU is involved.  Huge pages would probably help in this case, but I'm not sure if there is any way to guarantee that we get huge pages for system memory.  I guess we could use CMA or something like that.

Alex

> 
> Regards,
> Lucas
> 
> > Cc: Alex Deucher <alexander.deucher@xxxxxxx>
> > Bug:
> > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitl
> >
> ab.freedesktop.org%2Fdrm%2Famd%2Fissues%2F961&amp;data=02%7C01%
> 7Calexa
> >
> nder.deucher%40amd.com%7C30540b2bf2be417c4d9508d7765bf07f%7C3dd
> 8961fe4
> >
> 884e608e11a82d994e183d%7C0%7C0%7C637108010075463266&amp;sdata=1
> ZIZUWos
> > cPiB4auOY10jlGzoFeWszYMDBQG0CtrrOO8%3D&amp;reserved=0
> > Signed-off-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
> > ---
> > v2:
> > - Find Stoney graphics instead of host bridge.
> >
> >  drivers/iommu/amd_iommu_init.c | 13 ++++++++++++-
> >  1 file changed, 12 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/iommu/amd_iommu_init.c
> > b/drivers/iommu/amd_iommu_init.c index 568c52317757..139aa6fdadda
> > 100644
> > --- a/drivers/iommu/amd_iommu_init.c
> > +++ b/drivers/iommu/amd_iommu_init.c
> > @@ -2516,6 +2516,7 @@ static int __init early_amd_iommu_init(void)
> >  	struct acpi_table_header *ivrs_base;
> >  	acpi_status status;
> >  	int i, remap_cache_sz, ret = 0;
> > +	u32 pci_id;
> >
> >  	if (!amd_iommu_detected)
> >  		return -ENODEV;
> > @@ -2603,6 +2604,16 @@ static int __init early_amd_iommu_init(void)
> >  	if (ret)
> >  		goto out;
> >
> > +	/* Disable IOMMU if there's Stoney Ridge graphics */
> > +	for (i = 0; i < 32; i++) {
> > +		pci_id = read_pci_config(0, i, 0, 0);
> > +		if ((pci_id & 0xffff) == 0x1002 && (pci_id >> 16) == 0x98e4) {
> > +			pr_info("Disable IOMMU on Stoney Ridge\n");
> > +			amd_iommu_disabled = true;
> > +			break;
> > +		}
> > +	}
> > +
> >  	/* Disable any previously enabled IOMMUs */
> >  	if (!is_kdump_kernel() || amd_iommu_disabled)
> >  		disable_iommus();
> > @@ -2711,7 +2722,7 @@ static int __init state_next(void)
> >  		ret = early_amd_iommu_init();
> >  		init_state = ret ? IOMMU_INIT_ERROR :
> IOMMU_ACPI_FINISHED;
> >  		if (init_state == IOMMU_ACPI_FINISHED &&
> amd_iommu_disabled) {
> > -			pr_info("AMD IOMMU disabled on kernel command-
> line\n");
> > +			pr_info("AMD IOMMU disabled\n");
> >  			init_state = IOMMU_CMDLINE_DISABLED;
> >  			ret = -EINVAL;
> >  		}