On Mon, Nov 04, 2019 at 07:48:35PM +0100, Thomas Zimmermann wrote:
> Hi Daniel
>
> Am 04.11.19 um 10:55 schrieb Daniel Vetter:
> > On Mon, Oct 28, 2019 at 09:13:47AM +0100, Thomas Zimmermann wrote:
> >> Hi
> >>
> >> Am 25.10.19 um 20:54 schrieb Daniel Vetter:
> >>> On Fri, Oct 25, 2019 at 7:26 PM Thomas Zimmermann <tzimmermann@xxxxxxx> wrote:
> >>>>
> >>>> Hi
> >>>>
> >>>> Am 25.10.19 um 17:46 schrieb Noralf Trønnes:
> >>>>>
> >>>>>
> >>>>> Den 25.10.2019 11.27, skrev Thomas Zimmermann:
> >>>>>> There are no users of drm_fb_helper_defio_init(), so we can remove
> >>>>>> it. The documenation around defio support is a bit misleading and
> >>>>>> should mention compatibility issues with SHMEM helpers. Clarify this.
> >>>>>>
> >>>>>> Signed-off-by: Thomas Zimmermann <tzimmermann@xxxxxxx>
> >>>>>> ---
> >>>>>> drivers/gpu/drm/drm_fb_helper.c | 61 +++++++--------------------------
> >>>>>> include/drm/drm_fb_helper.h | 1 -
> >>>>>> 2 files changed, 13 insertions(+), 49 deletions(-)
> >>>>>>
> >>>>>> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> >>>>>> index b75ae8555baf..8ebeccdeed23 100644
> >>>>>> --- a/drivers/gpu/drm/drm_fb_helper.c
> >>>>>> +++ b/drivers/gpu/drm/drm_fb_helper.c
> >>>>>> @@ -92,9 +92,12 @@ static DEFINE_MUTEX(kernel_fb_helper_lock);
> >>>>>> *
> >>>>>> * Drivers that support a dumb buffer with a virtual address and mmap support,
> >>>>>> * should try out the generic fbdev emulation using drm_fbdev_generic_setup().
> >>>>>> + * It will automatically set up deferred I/O if the driver requires a shadow
> >>>>>> + * buffer.
> >>>>>> *
> >>>>>> - * Setup fbdev emulation by calling drm_fb_helper_fbdev_setup() and tear it
> >>>>>> - * down by calling drm_fb_helper_fbdev_teardown().
> >>>>>> + * For other drivers, setup fbdev emulation by calling
> >>>>>> + * drm_fb_helper_fbdev_setup() and tear it down by calling
> >>>>>> + * drm_fb_helper_fbdev_teardown().
> >>>>>> *
> >>>>>> * At runtime drivers should restore the fbdev console by using
> >>>>>> * drm_fb_helper_lastclose() as their &drm_driver.lastclose callback.
> >>>>>> @@ -127,8 +130,10 @@ static DEFINE_MUTEX(kernel_fb_helper_lock);
> >>>>>> * always run in process context since the fb_*() function could be running in
> >>>>>> * atomic context. If drm_fb_helper_deferred_io() is used as the deferred_io
> >>>>>> * callback it will also schedule dirty_work with the damage collected from the
> >>>>>> - * mmap page writes. Drivers can use drm_fb_helper_defio_init() to setup
> >>>>>> - * deferred I/O (coupled with drm_fb_helper_fbdev_teardown()).
> >>>>>> + * mmap page writes.
> >>>>>> + *
> >>>>>> + * Deferred I/O is not compatible with SHMEM. Such drivers should request an
> >>>>>> + * fbdev shadow buffer and call drm_fbdev_generic_setup() instead.
> >>>>>> */
> >>>>>>
> >>>>>> static void drm_fb_helper_restore_lut_atomic(struct drm_crtc *crtc)
> >>>>>> @@ -679,49 +684,6 @@ void drm_fb_helper_deferred_io(struct fb_info *info,
> >>>>>> }
> >>>>>> EXPORT_SYMBOL(drm_fb_helper_deferred_io);
> >>>>>>
> >>>>>> -/**
> >>>>>> - * drm_fb_helper_defio_init - fbdev deferred I/O initialization
> >>>>>> - * @fb_helper: driver-allocated fbdev helper
> >>>>>> - *
> >>>>>> - * This function allocates &fb_deferred_io, sets callback to
> >>>>>> - * drm_fb_helper_deferred_io(), delay to 50ms and calls fb_deferred_io_init().
> >>>>>> - * It should be called from the &drm_fb_helper_funcs->fb_probe callback.
> >>>>>> - * drm_fb_helper_fbdev_teardown() cleans up deferred I/O.
> >>>>>> - *
> >>>>>> - * NOTE: A copy of &fb_ops is made and assigned to &info->fbops. This is done
> >>>>>> - * because fb_deferred_io_cleanup() clears &fbops->fb_mmap and would thereby
> >>>>>> - * affect other instances of that &fb_ops.
> >>>>>> - *
> >>>>>> - * Returns:
> >>>>>> - * 0 on success or a negative error code on failure.
> >>>>>> - */
> >>>>>> -int drm_fb_helper_defio_init(struct drm_fb_helper *fb_helper)
> >>>>>> -{
> >>>>>> - struct fb_info *info = fb_helper->fbdev;
> >>>>>> - struct fb_deferred_io *fbdefio;
> >>>>>> - struct fb_ops *fbops;
> >>>>>> -
> >>>>>> - fbdefio = kzalloc(sizeof(*fbdefio), GFP_KERNEL);
> >>>>>> - fbops = kzalloc(sizeof(*fbops), GFP_KERNEL);
> >>>>>> - if (!fbdefio || !fbops) {
> >>>>>> - kfree(fbdefio);
> >>>>>> - kfree(fbops);
> >>>>>> - return -ENOMEM;
> >>>>>> - }
> >>>>>> -
> >>>>>> - info->fbdefio = fbdefio;
> >>>>>> - fbdefio->delay = msecs_to_jiffies(50);
> >>>>>> - fbdefio->deferred_io = drm_fb_helper_deferred_io;
> >>>>>> -
> >>>>>> - *fbops = *info->fbops;
> >>>>>> - info->fbops = fbops;
> >>>>>> -
> >>>>>> - fb_deferred_io_init(info);
> >>>>>> -
> >>>>>> - return 0;
> >>>>>> -}
> >>>>>> -EXPORT_SYMBOL(drm_fb_helper_defio_init);
> >>>>>> -
> >>>>>
> >>>>> With this gone you can remove the defio cleanup part from
> >>>>> drm_fb_helper_fbdev_teardown() as well.
> >>>>>
> >>>>> And I see that there's no users left of that function (the same with
> >>>>> *_seup()). Would be nice if you just removed them. I made them before I
> >>>>> embarked on the generic fbdev solution. I think it's better to try and
> >>>>> make the generic emulation usable for "everyone" and avoid the need for
> >>>>> drivers to have to do their own special stuff.
> >>>>
> >>>> The last user was vboxvideo, which I converted ~2 weeks ago. I haven't
> >>>> removed them yet, as there's a TODO item to convert drivers over to them.
> >>>>
> >>>> From a quick 'git grep':
> >>>>
> >>>> Most drivers that uses drm_fb_helper_sys_*() in its fb_ops could
> >>>> probably be converted over to generic fbdev. That's hibmc (I have some
> >>>> untested patches for it), msm, omapdrm, tegra, and udl (currently being
> >>>> converted).
> >>>>
> >>>> Only nouveau and gma500 have some form of HW acceleration.
> >>>>
> >>>> The rest of the drivers (10) uses drm_fb_helper_cfb_*() functions. Are
> >>>> these strictly required, or can they be made available within generic fbdev?
> >>>
> >>> Take a pile of digging through a few fb horrors, but this boils down to:
> >>>
> >>> *sys* functions operate on an fb that works like normal system memory.
> >>>
> >>> *cfb* functions operate on void __iomem * instead. Which makes a huge
> >>> difference on some architectures, but afaiui neither x86 nor arm care.
> >>> On ppc it seems to make a difference sometimes.
> >>>
> >>> So for shmem we should use *sys* functions, for vram *cfb*.
> >>>
> >>> And that leads me to realizing that drm_gem_vram_vmap has the totally
> >>> wrong type, it should be void __iomem *. There's this fancy is_iomem
> >>> parameter for ttm_kmap_obj_virtual that should help you figure out the
> >>> right type, but only nouveau bothers to implement this correctly.
> >>>
> >>> I'm honestly not sure whether we should care.
> >>
> >> I remember getting an eamil about this from some automated test system.
> >> I haven't had time to change it and it's apparently not urgent.
> >>
> >> If we really want to fix the problem, we'd have to propagage is_iomem
> >> through the DRM core; probably touching every vmap callback. OTOH this
> >> might not be a bad thing. With is_iomem available in the generic fbdev
> >> emulation, it could select between sys and cfb functions and support
> >> drivers with cfb-based fbdev as well.
> >
> > is_iomem doesn't work, it's hack, since the change is in the function
> > prototypes/signatures. So we'd need to make sure we have 2 sets of kernel
> > mapping functions for everything, plus 2 sets of anything that accesses
> > through the kernel.
> >
> > It's a huge task, and I'm really not sure we have any architecture we care
> > about this ...
>
> But the caller of vmap() doesn't know if it is I/O memory. It has to get
> this information via vmap() from the memory manager. Having two
> individual vmap() functions (and picking the correct one) would be quite
> a burden to the caller.
>
> But what's wrong with casting the returned address to void __iomem* if
> is_iomem is true?
>
> To get an idea of how well this works, I made a patchset to propagate
> is_iomem through all the vmap() interfaces in DRM. I found that most
> drivers' memory managers are SHMEM- or CMA-based and don't have to
> bother. The rest is based on TTM, which already returns the correct
> value for is_iomem. I didn't modify dma-buf interfaces and simply
> assumed 'false' here.
>
> The final patches modify the fbdev emulation to use either sys or cfb
> functions, depending on the value of is_iomem. Admittedly, I don't have
> the hardware where it makes a differences, but the change wasn't that
> hard either.
>
> I can probably post the patchset later this week for RFC.
The big reason behind __iomem is that you can use sparse to make sure you
got it right. With is_iomem and the explicit cast, you'll lose that
information. Which means no one will get it right (viz entire current drm
except nouveau).
That's why I think the 2 paths would be a lot nicer, and callers would
need to first call the system memory mmap, then the iomem mmap, until they
have a non-NULL pointer. Since they need the duplicated code anyway.
Other option would be we do an entire new pointer like the below:
struct opaque_buffer_ptr {
union { void * smem; void * __iomem iomem ; };
bool is_iomem;
};
And then an entire new set of functions that deals in this new primitive.
But unloading the is_iomem explicitly on drivers, expecting them to get it
right without the help or sparse, seems futile.
Also, all of this is _huuuuuuuuge_ amounts of work, and I'm still not sure
where we need it.
-Daniel
--
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
