Linus Walleij <linus.walleij@xxxxxxxxxx> writes:
> On Sun, Apr 8, 2018 at 3:08 AM, Eric Anholt <eric@xxxxxxxxxx> wrote:
>
>>> if (of_property_read_u32(dev->of_node, "max-memory-bandwidth",
>>> &priv->memory_bw)) {
>>> dev_info(dev, "no max memory bandwidth specified, assume unlimited\n");
>>> @@ -275,7 +280,8 @@ static int pl111_amba_probe(struct amba_device *amba_dev,
>>> priv->regs = devm_ioremap_resource(dev, &amba_dev->res);
>>> if (IS_ERR(priv->regs)) {
>>> dev_err(dev, "%s failed mmio\n", __func__);
>>> - return PTR_ERR(priv->regs);
>>> + ret = PTR_ERR(priv->regs);
>>> + goto mem_rel;
>>> }
>>
>> Shouldn't this error path be jumping to dev_unref, as well? We're
>> trying to match drm_dev_alloc(), right?
>
> OK I fixed.
>
>> I'm still a little bothered that we're allowing imports to pl111 of CMA
>> buffers that we can't scan out. Could we just refuse all
>> .gem_prime_import*() on this platform?
>
> I am sorry but I do not understand how CMA, dmabuf and GEM play
> together to decode this and figure out what to do.
>
> Do you mean that if we find device assigned memory, i.e. that there
> is this special memory which is all the controller can scan out,
> I should just implement .gem_prime_impport() instead of the
> currently assigned drm_gem_prime_import() to something just
> returning return ERR_PTR(-EINVAL); so it always fails?
>
> What about .gem_prime_import_sg_table()? Exporting should
> work fine since the memory is always readable by CPU.
Oh, I think you still want drm_gem_prime_import()'s path for "I'm
importing an fd from this same driver to into another process". So,
yeah, have our .import_sg_table hook throw -EINVAL if called on the
device memory platform, and otherwise call down to
drm_gem_cma_prime_import_sg_table().
Attachment:
signature.asc
Description: PGP signature
_______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
