Le lundi 22 octobre 2018 à 12:37 +0900, Tomasz Figa a écrit :
> Hi Philipp,
> 
> On Mon, Oct 22, 2018 at 1:28 AM Philipp Zabel <pza@pengutronix.de> wrote:
> > 
> > On Wed, Oct 03, 2018 at 05:24:39PM +0900, Tomasz Figa wrote:
> > [...]
> > > > Yes, but that would fall in a complete redesign I guess. The buffer
> > > > allocation scheme is very inflexible. You can't have buffers of two
> > > > dimensions allocated at the same time for the same queue. Worst, you
> > > > cannot leave even 1 buffer as your scannout buffer while reallocating
> > > > new buffers, this is not permitted by the framework (in software). As a
> > > > side effect, there is no way to optimize the resolution changes, you
> > > > even have to copy your scannout buffer on the CPU, to free it in order
> > > > to proceed. Resolution changes are thus painfully slow, by design.
> > 
> > [...]
> > > Also, I fail to understand the scanout issue. If one exports a vb2
> > > buffer to a DMA-buf and import it to the scanout engine, it can keep
> > > scanning out from it as long as it want, because the DMA-buf will hold
> > > a reference on the buffer, even if it's removed from the vb2 queue.
> > 
> > REQBUFS 0 fails if the vb2 buffer is still in use, including from dmabuf
> > attachments: vb2_buffer_in_use checks the num_users memop. The refcount
> > returned by num_users shared between the vmarea handler and dmabuf ops,
> > so any dmabuf attachment counts towards in_use.
> 
> Ah, right. I've managed to completely forget about it, since we have a
> downstream patch that we attempted to upstream earlier [1], but didn't
> have a chance to follow up on the comments and there wasn't much
> interest in it in general.
> 
> [1] https://lore.kernel.org/patchwork/patch/607853/
> 
> Perhaps it would be worth reviving?

In this patch we should consider a way to tell userspace that this has
been opt in, otherwise existing userspace will have to remain using
sub-optimal copy based reclaiming in order to ensure that renegotiation
can work on older kernel tool. At worst someone could probably do trial
and error (reqbufs(1)/mmap/reqbufs(0)) but on CMA with large buffers
this introduces extra startup time.

> 
> Best regards,
> Tomasz