[Intel-gfx] [PATCH] drm/i915: Convert WARNs during userptr revoke to SIGBUS

[Dave Gordon]

On 09/10/15 18:26, Chris Wilson wrote:
> On Fri, Oct 09, 2015 at 07:14:02PM +0200, Daniel Vetter wrote:
>> On Fri, Oct 09, 2015 at 10:03:14AM +0100, Tvrtko Ursulin wrote:
>>>
>>> On 09/10/15 09:55, Daniel Vetter wrote:
>>>> On Fri, Oct 09, 2015 at 09:40:53AM +0100, Chris Wilson wrote:
>>>>> On Fri, Oct 09, 2015 at 09:48:01AM +0200, Daniel Vetter wrote:
>>>>>> On Thu, Oct 08, 2015 at 10:45:47AM +0100, Tvrtko Ursulin wrote:
>>>>>> The concern is that this isn't how SIG_SEGV works, it's a signal the
>>>>>> thread who made the invalid access gets directly. You never get a SIG_SEGV
>>>>>> for bad access someone else has made. So essentially it's new ABI.
>>>>>
>>>>> SIGBUS. For which the answer is yes, you can and do get SIGBUS for
>>>>> actions taken by other processes.
>>>>
>>>> Oh right I always forget that SIGBUS aliases with SIGIO. Anyway if
>>>> userspace wants SIGIO we just need to provide it with a pollable fd and
>>>> then it can use fcntl to make that happen. That's imo a much better api
>>>> than unconditionally throwing around signals. Also we already have the
>>>> reset stats ioctl to tell userspace that its gpu context is toats. If
>>>> anyone wants that to be pollable (or even send SIGIO) I think we should
>>>> extend that, with all the usual "needs userspace&igt" stuff on top.
>>>
>>> I don't see that this notification can be optional. Process is confused
>>> about its memory map use so should die. :)
>>>
>>> This is not a GPU error/hang - this is the process doing stupid things.
>>>
>>> MMU notifiers do not support decision making otherwise we could say
>>> -ETXTBUSY or something on munmap, but we can't. Not even sure that it would
>>> help in all cases, would have to fail clone as well and who knows what.
>>
>> So what happens if the gpu just keeps using the memory? It'll all be
>> horribly undefined behaviour and eventually it'll die on an -EFAULT in
>> execbuf, but does anything else bad happen?
>
> We don't see an EFAULT unless a miracle occurs, and the stale pages
> continue to be read/written by other processes (as well as the client).
> Horribly undefined behaviour with a misinformation leak.
> -Chris

I think SIGBUS would be a good notification. It's the sort of outcome 
you expect when a privileged thread on the CPU or any sort of DMA-master 
device incurs an access fault on physical memory or I/O mapped register 
space. One explanation I found suggests:

	Another reason where SIGBUS can generate is explained below:

	You are currently using a external I/O device by mapping the
	device memory mapping into the system memory (Memory mapped
	I/O). You have used it. And now, you have disconnected it
	gracefully. But, somehow your code is trying to use an
	previously used address still in your code. The result in this
	case will be an SIGBUS, the reason is BUS_ADRERR, "non-existent
	physical address".

See http://cquestion.blogspot.com/2008/03/sigbus-vs-sigsegv.html

In this case, (we assume that) the GPU is going to continue to access 
the "physical" (PPGTT?) address of the (virtual) memory that the process 
is trying to revoke its access to. And while it might make sense to 
remove a buffer from the CPU's mapping while the GPU was still accessing 
it, it really makes no sense to delete a GTT mapping that the GPU may 
still (asynchronously) be accessing. So either we have to kill the 
process's outstanding tasks on the GPU (context-specific reset?) or fail 
the unmap (and shoot the process for trying to sabotage the GPU?).

Or ... could we decouple the pages? Duplicate them as for copy-on-write, 
and give one copy to the user process and the other to the GPU?
Of course the actual content of the page might be indeterminate if the 
GPU were writing it while the CPU was taking a copy ... does this make 
any sense?

.Dave.