Date:   Wed, 7 Jul 2021 17:06:35 +0800
From:   Kai-Heng Feng <kai.heng.feng@...onical.com>
To:     Robin Murphy <robin.murphy@....com>
Cc:     Joerg Roedel <joro@...tes.org>, will@...nel.org,
        "open list:IOMMU DRIVERS" <iommu@...ts.linux-foundation.org>,
        open list <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] iommu: Fallback to default setting when def_domain_type()
 callback returns 0

On Wed, Jul 7, 2021 at 2:03 AM Robin Murphy <robin.murphy@....com> wrote:
>
> On 2021-07-06 17:21, Kai-Heng Feng wrote:
> > On Tue, Jul 6, 2021 at 5:27 PM Robin Murphy <robin.murphy@....com> wrote:
> >>
> >> On 2021-07-06 07:51, Kai-Heng Feng wrote:
> >>> Commit 28b41e2c6aeb ("iommu: Move def_domain type check for untrusted
> >>> device into core") not only moved the check for untrusted device to
> >>> IOMMU core, it also introduced a behavioral change by returning
> >>> def_domain_type() directly without checking its return value. That makes
> >>> many devices no longer use the default IOMMU setting.
> >>>
> >>> So revert back to the old behavior which defaults to
> >>> iommu_def_domain_type when driver callback returns 0.
> >>>
> >>> Fixes: 28b41e2c6aeb ("iommu: Move def_domain type check for untrusted device into core")
> >>
> >> Are you sure about that? From that same commit:
> >>
> >> @@ -1507,7 +1509,7 @@ static int iommu_alloc_default_domain(struct
> >> iommu_group *group,
> >>           if (group->default_domain)
> >>                   return 0;
> >>
> >> -       type = iommu_get_def_domain_type(dev);
> >> +       type = iommu_get_def_domain_type(dev) ? : iommu_def_domain_type;
> >>
> >>           return iommu_group_alloc_default_domain(dev->bus, group, type);
> >>    }
> >>
> >> AFAICS the other two callers should also handle 0 correctly. Have you
> >> seen a problem in practice?
> >
> > Thanks for pointing out how the return value is being handled by the callers.
> > However, the same check is missing in probe_get_default_domain_type():
> > static int probe_get_default_domain_type(struct device *dev, void *data)
> > {
> >          struct __group_domain_type *gtype = data;
> >          unsigned int type = iommu_get_def_domain_type(dev);
> > ...
> > }
>
> I'm still not following - the next line right after that is "if (type)",
> which means it won't touch gtype, and if that happens for every
> iteration, probe_alloc_default_domain() subsequently hits its "if
> (!gtype.type)" condition and still ends up with iommu_def_domain_type.
> This *was* one of the other two callers I was talking about (the second
> being iommu_change_dev_def_domain()), and in fact on second look I think
> your proposed change will actually break this logic, since it's
> necessary to differentiate between a specific type being requested for
> the given device, and a "don't care" response which only implies to use
> the global default type if it's still standing after *all* the
> appropriate devices have been queried.

You are right, what I am seeing here is that the AMD GFX is using
IOMMU_DOMAIN_IDENTITY, and makes other devices in the same group,
including the Realtek WiFi, to use IOMMU_DOMAIN_IDENTITY as well.

>
> > I personally prefer the old way instead of open coding with ternary
> > operator, so I'll do that in v2.
> >
> > In practice, this causes a kernel panic when probing Realtek WiFi.
> > Because of the bug, dma_ops isn't set by probe_finalize(),
> > dma_map_single() falls back to swiotlb which isn't set and caused a
> > kernel panic.
>
> Hmm, but if that's the case, wouldn't it still be a problem anyway if
> the end result was IOMMU_DOMAIN_IDENTITY? I can't claim to fully
> understand the x86 swiotlb and no_iommu dance, but nothing really stands
> out to give me confidence that it handles the passthrough options correctly.

Yes, I think AMD IOMMU driver needs more thourough check on
static void __init amd_iommu_init_dma_ops(void)
{
        swiotlb = (iommu_default_passthrough() || sme_me_mask) ? 1 : 0;
...
}

Since swiotlb gets disabled by the check, but the Realtek WiFi is only
capable of doing 32bit DMA, the kernel panics because
io_tlb_default_mem is NULL.

Thanks again for pointing to the right direction, this patch is indeed
incorrect.

Kai-Heng

>
> Robin.
>
> > I didn't attach the panic log because the system simply is frozen at
> > that point so the message is not logged to the storage.
> > I'll see if I can find another way to collect the log and attach it in v2.
> >
> > Kai-Heng
> >
> >>
> >> Robin.
> >>
> >>> Signed-off-by: Kai-Heng Feng <kai.heng.feng@...onical.com>
> >>> ---
> >>>    drivers/iommu/iommu.c | 5 +++--
> >>>    1 file changed, 3 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> >>> index 5419c4b9f27a..faac4f795025 100644
> >>> --- a/drivers/iommu/iommu.c
> >>> +++ b/drivers/iommu/iommu.c
> >>> @@ -1507,14 +1507,15 @@ EXPORT_SYMBOL_GPL(fsl_mc_device_group);
> >>>    static int iommu_get_def_domain_type(struct device *dev)
> >>>    {
> >>>        const struct iommu_ops *ops = dev->bus->iommu_ops;
> >>> +     unsigned int type = 0;
> >>>
> >>>        if (dev_is_pci(dev) && to_pci_dev(dev)->untrusted)
> >>>                return IOMMU_DOMAIN_DMA;
> >>>
> >>>        if (ops->def_domain_type)
> >>> -             return ops->def_domain_type(dev);
> >>> +             type = ops->def_domain_type(dev);
> >>>
> >>> -     return 0;
> >>> +     return (type == 0) ? iommu_def_domain_type : type;
> >>>    }
> >>>
> >>>    static int iommu_group_alloc_default_domain(struct bus_type *bus,
> >>>

Powered by blists - more mailing lists