Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Daniel Drake <drake@endlessm.com>
To:     joro@8bytes.org
Cc:     agross@kernel.org, baolu.lu@linux.intel.com,
        bjorn.andersson@linaro.org, dwmw2@infradead.org,
        gerald.schaefer@de.ibm.com, heiko@sntech.de,
        iommu@lists.linux-foundation.org, jean-philippe@linaro.org,
        jonathanh@nvidia.com, kgene@kernel.org, krzk@kernel.org,
        linux-arm-msm@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-mediatek@lists.infradead.org,
        linux-rockchip@lists.infradead.org, linux-s390@vger.kernel.org,
        linux-samsung-soc@vger.kernel.org, linux-tegra@vger.kernel.org,
        m.szyprowski@samsung.com, matthias.bgg@gmail.com,
        robdclark@gmail.com, robin.murphy@arm.com,
        thierry.reding@gmail.com,
        virtualization@lists.linux-foundation.org, will@kernel.org,
        jonathan.derrick@intel.com
Subject: Re: [PATCH v2 00/33] iommu: Move iommu_group setup to IOMMU core code
Date:   Fri, 17 Apr 2020 09:03:35 +0800
Message-Id: <20200417010335.31739-1-drake@endlessm.com>
In-Reply-To: <20200414131542.25608-1-joro@8bytes.org>
References: <20200414131542.25608-1-joro@8bytes.org>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Hi Joerg,

> Hi,
> 
> here is the second version of this patch-set. The first version with
> some more introductory text can be found here:
> 
> 	https://lore.kernel.org/lkml/20200407183742.4344-1-joro@8bytes.org/

Thanks for the continued improvements in this area!

I may have spotted a problem with setups like VMD.

The core PCI bus is set up during early boot.
Then, for the PCI bus, we reach iommu_bus_init() -> bus_iommu_probe().
In there, we call probe_iommu_group() -> dev_iommu_get() for each PCI
device, which allocates dev->iommu in each case. So far so good.

The problem is that this is the last time that we'll call dev_iommu_get().
If any PCI bus devices get added after this point, they do not get passed
to dev_iommu_get().

So when the vmd module gets loaded later, and creates more PCI devices,
we end up in iommu_bus_notifier() -> iommu_probe_device()
-> __iommu_probe_device() which does:

	dev->iommu->iommu_dev = iommu_dev;

dev->iommu-> is a NULL dereference because dev_iommu_get() was never
called for this new device.

Daniel