On Thu, Sep 13, 2018 at 04:26:53PM +0200, Neil Armstrong wrote: > Hi Daniel, > > On 13/09/2018 15:21, Daniel Vetter wrote: > > On Wed, Sep 12, 2018 at 01:06:07PM +0200, Noralf Trønnes wrote: > >> > >> Den 12.09.2018 12.57, skrev Noralf Trønnes: > >>> (Cc: Daniel Vetter) > >>> > >> > >> Somehow that CC was dropped somewhere after leaving email client. > >> Trying once more. > > > > Yeah I just made myself unpopular. If you want SMEM_START, then you need > > to carry a local patch now. virt_to_pfn on the vmap should work. It's > > about 2 lines, including the change to drop HIDE_SMEM_START. > > Would it be totally unacceptable to add such 2 line : > - enabled by a specific config depending on EXPERT and narrowed to specific platforms > - printing a big fat pr_warning_once when used > - with a big fat comment specifying when this code will be dropped and why we should not activate it module_param_debug auto-taints your kernel, I'd just go with that. Plus CONFIG_EXPERT (or CONFIG_BROKEN). But yes, I think that's something I'll happily ack. Must be acked by Dave (and perhaps a few others too), since defacto this means we now suddenly care about closed source blobs. I'd say get someone from amd and a few of the driver maintainers on board with this (nouveau, Eric, Rob, Lucas, ...), to make sure it really represents community consensus. Cheers, Daniel > > Neil > > > > And if consensus is that hiding SMEM_START is not a nice idea, then I'm > > sure we can reintroduce it through some slightly unpretty backdoor, even > > with Noralf's generic code. So not really a good reason to reject the > > cleanup I think. > > -Daniel > > > > > >>> Den 12.09.2018 11.56, skrev Maxime Ripard: > >>>> On Wed, Sep 12, 2018 at 11:48:34AM +0200, Neil Armstrong wrote: > >>>>> Hi Noralf, > >>>>> > >>>>> On 08/09/2018 15:46, Noralf Trønnes wrote: > >>>>>> The CMA helper is already using the > >>>>>> drm_fb_helper_generic_probe part of > >>>>>> the generic fbdev emulation. This patch makes full use of the generic > >>>>>> fbdev emulation by using its drm_client callbacks. This means that > >>>>>> drm_mode_config_funcs->output_poll_changed and > >>>>>> drm_driver->lastclose are > >>>>>> now handled by the emulation code. Additionally fbdev > >>>>>> unregister happens > >>>>>> automatically on drm_dev_unregister(). > >>>>>> > >>>>>> The drm_fbdev_generic_setup() call is put after > >>>>>> drm_dev_register() in the > >>>>>> driver. This is done to highlight the fact that fbdev emulation is an > >>>>>> internal client that makes use of the driver, it is not part of the > >>>>>> driver as such. If fbdev setup fails, an error is printed, > >>>>>> but the driver > >>>>>> succeeds probing. > >>>>> I can't really ack/nack this move, on one side it's great to have a > >>>>> generic fbdev emulation instead of the old fbdev code, but on the > >>>>> other side the Amlogic platform (like allwinner, samsung, rockchip, > >>>>> ...) still relies on an Fbdev variant of the libMali that uses > >>>>> smem_start... > >>>>> > >>>>> I know it's dirty and fbdev based code should be legacy now, but I > >>>>> consider it like an user-space breakage... > >>>>> > >>>>> I'll be happy if ARM provided it's Mali vendors a Fbdev libMali that > >>>>> could use FBINFO_HIDE_SMEM_START and allocate BO's from the DRM > >>>>> driver, it won't be the case and it will never be the case until the > >>>>> Lima projects finalizes. > >>>>> > >>>>> So for me it's a no-go until Lima lands upstream in Linux *and* Mesa. > >>>> My feelings exactly. If the choice is between reducing the DRM driver > >>>> by a couple of dozens of lines or keeping the mali working, I'll pick > >>>> the latter, every single day. > >>> > >>> I don't know the reasoning for blocking smem_start other than what Daniel > >>> wrote in these commit messages: > >>> > >>> da6c7707caf3736c1cf968606bd97c07e79625d4 > >>> fbdev: Add FBINFO_HIDE_SMEM_START flag > >>> > >>> DRM drivers really, really, really don't want random userspace to > >>> share buffer behind it's back, bypassing the dma-buf buffer sharing > >>> machanism. For that reason we've ruthlessly rejected any IOCTL > >>> exposing the physical address of any graphics buffer. > >>> > >>> Unfortunately fbdev comes with that built-in. We could just set > >>> smem_start to 0, but that means we'd have to hand-roll our own fb_mmap > >>> implementation. For good reasons many drivers do that, but > >>> smem_start/length is still super convenient. > >>> > >>> Hence instead just stop the leak in the ioctl, to keep fb mmap working > >>> as-is. A second patch will set this flag for all drm drivers. > >>> > >>> > >>> 6be8f3bd2c78915a9f3a058a346ae93068d35c01 > >>> drm/fb: Stop leaking physical address > >>> > >>> For buffer sharing, use dma-buf instead. We can't set smem_start to 0 > >>> unconditionally since that's used by the fbdev mmap default > >>> implementation. And we have plenty of userspace which would like to > >>> keep that working. > >>> > >>> This might break legit userspace - if it does we need to look at a > >>> case-by-cases basis how to handle that. Worst case I expect overrides > >>> for only specific drivers, since anything remotely modern should be > >>> using dma-buf/prime now (which is about 7 years old now for DRM > >>> drivers). > >>> > >>> This issue was uncovered because Noralf's rework to implement a > >>> generic fb_probe also implements it's own fb_mmap callback. Which > >>> means smem_start didn't have to be set anymore, which blew up some > >>> blob in userspace rather badly. > >>> > >>> > >>> Noralf. > >>> > >>> _______________________________________________ > >>> dri-devel mailing list > >>> dri-devel@xxxxxxxxxxxxxxxxxxxxx > >>> https://lists.freedesktop.org/mailman/listinfo/dri-devel > >>> > >> > > > -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
