[PATCH 13/26] RFC drm/xe/eudebug: userptr vm pread/pwrite

[Christian König]

Am 29.01.25 um 19:18 schrieb Joonas Lahtinen:
>>> Would be great to reach a consensus on the high level details before
>>> spinning off further series addressing the smaller items.
>> I would say that attaching debug metadata to the GPU VMA doesn't look
>> like the best design, but if you just do that inside XE it won't affect
>> any other part of the kernel.
> It just grew out of convenience of implementation on the side of VM_BIND.
>
> The other alternative would be to maintain a secondary load map in the
> kernel in a separate data structure from GPU VMA.
>
> I was actually going to suggest such thing as a common DRM thing: GPU VMA
> metadata interface or parallel "GPU loadmap" interface. It'd allow for
> userspace tooling to more easier go from GPU EU IP to a module that
> was loaded at that address. Kind of a step 0 towards backtrace for GPU.
>
> Can you elaborate on what your concern is with the VMA metadata
> attachment?

In general we should try to avoid putting data into the kernel the 
kernel doesn't need.

In other words we don't put the debug metadata for the CPU process on 
the CPU VMA either.

You could reduce the amount of data massively if you just attach the 
source of the debug metadata to the GPU VMA.

E.g. instead of the symbol table just where to find it. A VA of the CPU 
process, a file system location or something like that.

>> My other concern I have is using ptrace_may_access, I would still try to
>> avoid that.
>>
>> What if you first grab the DRM render node file descriptor which
>> represents the GPU address space you want to debug with pidfd_getfd()
>> and then either create the eudebug file descriptor from an IOCTL or
>> implement the necessary IOCTLs on the DRM render node directly?
>>
>> That would make it unnecessary to export ptrace_may_access.
> We're prototyping this. At this point there are some caveats recognized:
>
> 1. There is a limitation that you don't get a notification when your
> target PID opens a DRM client, but GDB or other application would have
> to keep polling for the FDs. I'll have to check with the team how that
> would fit to GDB side.

Well there is the dnotify/inotify API which allows a process to be 
notified of certain filesystem events.

I never used it, but it potentially provides an event when a specific 
file is open.

On the other hand I have no idea what permissions are necessary to use 
it, potentially root.

> 2. Debugging multiple DRM clients (to same GPU) under one PID now
> requires separate debugger connections. This may break the way the debugger
> locking is currently implemented for the discovery phase to prevent parallel
> IOCTL from running. Will have to look into it once we have a working
> prototype.

Well multiple DRM clients would also have multiple GPU VM address 
spaces, wouldn't they?

So you would need multiple connections, one for each DRM client as well.

> 3. Last but not the least, we'll have to compare which LSM security
> modules and other conditions checked on the pidfd_getfd() paths for
> access restrictions.
>
> Reason for using ptrace_may_access() was to have a clear 1:1 mapping
> between user being allowed to ptrace() a PID to control CPU threads and
> do debugger IOCTL to control GPU threads. So if user can attach to a PID
> with GDB, they would certainly also be able to debug the GPU portion.
>
> If there is divergence, I don't see a big benefit in going to
> pidfd_getfd(). We're all the same not even exporting ptrace_may_access()
> and YOLO'ing the access check by comparing euid and such which is close
> to what ptrace does, but not exactly the same (just like pidfd_getfd()
> access checks).

Well that argumentation is exactly backward to why I suggested not using 
ptrace_may_access().

When an administrator used LSM to block pidfd_getfd() then a driver 
which uses a driver specific IOCTL to bypass that is actually a really 
bad idea.

> However I believe this would not be a fundamental blocker for the
> series? If this would be the only remaining disagreement, I guess we
> could just do CAP_ADMIN check initially before we can find something
> agreeable.

Well as soon as anything is merged upstream it becomes UAPI, which in 
turn means that changing it fundamentally becomes really hard to do.

What you could do is to expose the interface through debugfs and 
explicitly state that it isn't stable in any way.

Regards,
Christian.

>
> Regards, Joonas
>
>> Regards,
>> Christian.
>>
>>> Regards, Joonas
>>>
>>>> Regards,
>>>> Christian.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-xe/attachments/20250130/c30f085d/attachment.htm>