Re: [Linaro-mm-sig] Changing vma->vm_file in dma_buf_mmap()

From: Christian König
Date: Thu Sep 17 2020 - 03:11:51 EST

Am 17.09.20 um 08:23 schrieb Daniel Vetter:
On Wed, Sep 16, 2020 at 5:31 PM Christian König
<ckoenig.leichtzumerken@xxxxxxxxx> wrote:
Am 16.09.20 um 17:24 schrieb Daniel Vetter:
On Wed, Sep 16, 2020 at 4:14 PM Christian König
<christian.koenig@xxxxxxx> wrote:
Am 16.09.20 um 16:07 schrieb Jason Gunthorpe:
On Wed, Sep 16, 2020 at 11:53:59AM +0200, Daniel Vetter wrote:

But within the driver, we generally need thousands of these, and that
tends to bring fd exhaustion problems with it. That's why all the private
buffer objects which aren't shared with other process or other drivers are
handles only valid for a specific fd instance of the drm chardev (each
open gets their own namespace), and only for ioctls done on that chardev.
And for mmap we assign fake (but unique across all open fd on it) offsets
within the overall chardev. Hence all the pgoff mangling and re-mangling.
Are they still unique struct files? Just without a fdno?
Yes, exactly.
Not entirely, since dma-buf happened after drm chardev, so for that
historical reason the underlying struct file is shared, since it's the
drm chardev. But since that's per-device we don't have a problem in
practice with different vm_ops, since those are also per-device. But
yeah we could fish out some entirely hidden per-object struct file if
that's required for some mm internal reasons.
Hui? Ok that is just the handling in i915, isn't it?

As far as I know we create an unique struct file for each DMA-buf.
Yes dma-buf, but that gets forwarded to the original drm chardev which
originally exported the buffer. It's only there where the forwarding
chain stops. The other thing is that iirc we have a singleton
anon_inode behind all the dma-buf, so they'd share all the same
address_space and so would all alias for unmap_mapping_range (I think
at least).

Amdgpu works by using the address_space of the drm chardev into the struct file of DMA-buf instead.

I think that this is cleaner, but only by a little bit :)

Anyway I'm a bit concerned that we have so many different approaches for the same problem.





Hence why we'd like to be able to forward aliasing mappings and adjust the
file and pgoff, while hopefully everything keeps working. I thought this
would work, but Christian noticed it doesn't really.
It seems reasonable to me that the dma buf should be the owner of the
VMA, otherwise like you say, there is a big mess attaching the custom
vma ops and what not to the proper dma buf.

I don't see anything obviously against this in mmap_region() - why did
Chritian notice it doesn't really work?
To clarify I think this might work.

I just had the same "Is that legal?", "What about security?", etc..
questions you raised as well.

It seems like a source of trouble so I thought better ask somebody more
familiar with that.


dri-devel mailing list