Hi,

On Mon, Jul 25, 2016 at 5:39 PM, Matthias Brugger
[off-list ref] wrote:

On 20/07/16 05:01, Yong Wu wrote:

quoted

Currently the iommu consumer always call iommu_present to get whether
the iommu is ready. But in MTK IOMMU, this function can't indicate
this. The IOMMU call bus_set_iommu->mtk_iommu_add_device->
mtk_iommu_attach_device to parse the iommu data, then it's able to
transfer "struct mtk_smi_iommu" to SMI-LARB, and the iommu uses the
larbs as compoents, the iommu will finish its probe until all the larbs
probe done.

If the iommu consumer(like DRM) begin to probe after the time of
calling bus_set_iommu and before the time of SMI probe finish, it
will hang like this:

[    7.832359] Call trace:
[    7.834778] [<ffffffc000764424>] mtk_smi_larb_get+0x24/0xa8
[    7.840300] [<ffffffc0005a1390>] mtk_drm_crtc_enable+0x6c/0x450

Because the larb->mmu is NULL at that time.

In order to avoid this issue, we add a new interface
(mtk_smi_larb_is_ready) for checking whether the IOMMU and SMI have
finished their probe. If it return false, the iommu consumer should
probe-defer for the IOMMU and SMI.

Can't we just skip the functions in the probe and call bus_set_iommu only if
we were able to bind all components?
Something like this:

Note that we have to call bus_set_iommu() and actually have
.add_device() and .attach_device() called before any of the slave
devices probe. I found a similar problem with rockchip IOMMU after
adding power domain and runtime PM handling there. I also found that
current design of IOMMU core and related DMA mapping code is utterly
broken regarding the device add/probe ordering (no support for
deferring things properly).

So my idea is to keep .add_device() as is, since typically it doesn't
seem to require anything from the IOMMU hardware and just initializes
some per-device data, but make .attach_device() being able to defer
probe of that device if respective IOMMU has not probed yet. I'm still
in process of figuring out the right way to achieve it, though...

Best regards,
Tomasz