Re: [RFC v2 8/8] drm: tegra: Add gr2d device

[Thomas Hellstrom <thomas@shipmail.org>]

On 11/28/2012 02:33 PM, Lucas Stach wrote:
> Am Mittwoch, den 28.11.2012, 15:17 +0200 schrieb Terje Bergström:
>> On 28.11.2012 01:00, Dave Airlie wrote:
>>> We  generally aim for the first, to stop the gpu from reading/writing
>>> any memory it hasn't been granted access to,
>>> the second is nice to have though, but really requires a GPU with VM
>>> to implement properly.
>> I wonder if we should aim at root only access on Tegra20, and force
>> IOMMU on Tegra30 and fix the remaining issues we have with IOMMU. The
>> firewall turns out to be more complicated than I wished.
>>
>> Biggest problem is that we aim at zero-copy for everything possible,
>> including command streams. Kernel gets a handle to a command stream, but
>> the command stream is allocated by the user space process. So the user
>> space can tamper with the stream once it's been written to the host1x 2D
>> channel.
>>
> So this is obviously wrong. Userspace has to allocate a pushbuffer from
> the kernel just as every other buffer, then map it into it's own address
> space to push in commands. At submit time of the pushbuf kernel has to
> make sure that userspace is not able to access the memory any more, i.e.
> kernel shoots down the vma or pagetable of the vma.

To me this sounds very expensive. Zapping the page table requires a CPU 
TLB flush
on all cores that have touched the buffer, not to mention the kernel calls
required to set up the page table once the buffer is reused.

If this usage scheme then is combined with a command verifier or 
"firewall" that
reads from a *write-combined* pushbuffer performance will be bad. Really 
bad.

In such situations I think one should consider copy-from-user while 
validating, and
let user-space set up the command buffer in malloced memory.

/Thomas