Re: [PATCH] iommu: Avoid races around device probe

[Robin Murphy <robin.murphy@arm.com>]

On 2022-11-05 01:36, Brian Norris wrote:
> On Fri, Nov 04, 2022 at 07:51:43PM +0000, Robin Murphy wrote:
>> We currently have 3 different ways that __iommu_probe_device() may be
>> called, but no real guarantee that multiple callers can't tread on each
>> other, especially once asynchronous driver probe gets involved. It would
>> likely have taken a fair bit of luck to hit this previously, but commit
>> 57365a04c921 ("iommu: Move bus setup to IOMMU device registration") ups
>> the odds since now it's not just omap-iommu that may trigger multiple
>> bus_iommu_probe() calls in parallel if probing asynchronously.
>>
>> Add a lock to ensure we can't try to double-probe a device, and also
>> close some possible race windows to make sure we're truly robust against
>> trying to double-initialise a group via two different member devices.
>>
>> Reported-by: Brian Norris <briannorris@chromium.org>
>> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
>> ---
>>   drivers/iommu/iommu.c | 28 ++++++++++++++++++++++------
>>   1 file changed, 22 insertions(+), 6 deletions(-)
> 
> If I've tested appropriately (there's always room for operator error),
> this seems to resolve the problems I reported:
> 
> Tested-by: Brian Norris <briannorris@chromium.org>
> 
> I haven't reviewed closely enough to know how precisely this is a
> regression (your description sounds like you think the bug existed some
> time before that), but based on testing, this sounds like:
> 
> Fixes: 57365a04c921 ("iommu: Move bus setup to IOMMU device
> registration")

That commit did not introduce the race, just made it more visible. The 
underlying condition probably goes back at least 3 years to where we 
started allocating and freeing per-device data around what was then the 
ops->add_device() call.

In practice, you'd have to be absurdly lucky for an iommu_probe_device() 
call via {of,acpi}_dma_configure() to line up with bus_iommu_probe() 
touching the same device, but by inspection I think it's theoretically 
possible. Thus previously there was probably only a realistic chance of 
seeing it on certain OMAP systems, where the explicit bus_iommu_probe() 
calls could overlap if both instances probed in parallel - my commit 
just brings all the other drivers in line with that same behaviour via 
iommu_device_register(). Other systems - like Rockchip in particular - 
may have greater numbers of IOMMU instances and thus even more chance 
for parallel probes to line up just right.

Since nobody's ever reported real-world issues on OMAP (although it's 
quite likely nobody's ever tried driver_async_probe with omap-iommu 
anyway) there doesn't seem to be a compelling reason for backporting, so 
I didn't fancy spending hours digging through subsystem-wide history 
trying to figure out an appropriate fixes tag; as long as this can make 
6.1 that should be enough :)

Thanks,
Robin.

> But even if not, the report could probably use:
> 
> Link: https://lore.kernel.org/lkml/Y1CHh2oM5wyHs06J@google.com/
> 
> And most of all, thanks!
> 
> Brian