Den 13.04.2016 13:09, skrev Daniel Vetter:
On Fri, Apr 08, 2016 at 07:05:03PM +0200, Noralf Trønnes wrote:
This adds deferred io support if CONFIG_FB_DEFERRED_IO is enabled. Accumulated fbdev framebuffer changes are signaled using the callback (struct drm_framebuffer_funcs *)->dirty()
The drm_fb_helper_sys_*() functions will accumulate changes and schedule fb_info.deferred_work _if_ fb_info.fbdefio is set. This worker is used by the deferred io mmap code to signal that it has been collecting page faults. The page faults and/or other changes are then merged into a drm_clip_rect and passed to the framebuffer dirty() function.
The driver is responsible for setting up the fb_info.fbdefio structure and calling fb_deferred_io_init() using the provided callback: (struct fb_info *)->fbdefio->deferred_io = drm_fb_helper_deferred_io;
Signed-off-by: Noralf Trønnes noralf@tronnes.org
For this one it'd be awesome to throw patches for qxl/udl on top to remove their own hand-rolled implementations. Just to maximize the testing coverage of this new code. Should be doable to set up a qxl virtual machine quickly, but even without that we should be able to pull it in (since it's mostly just about removing code from these two drivers).
There are three fb_deferred_io users in drivers/gpu/drm: qxl, udl and vmwgfx.
drivers/gpu/drm/vmwgfx It doesn't use drm_fb_helper (uses the cfb_{fillrect,copyarea,imageblit} functions directly). This made me think that I should probably add fb_deferred_io support to drm_fb_helper_cfb_*() as well.
drivers/gpu/drm/udl First of all it has had deferred io disabled by default since 2013 (commit 677d23b). Secondly it handles damage directly in it's fb_{fillrect,copyarea,imageblit} functions, but defers to the next call if it is in atomic context. My patch always defers those function to a worker. The driver uses the drm_fb_helper_sys_* functions, so my patch would mess it up if someone would enable deferred io (module_param: fb_defio). So this driver isn't a good candidate for easy conversion also because it has different code paths for fbdev mmap damage and the other damages (although the code is similar). But it needs a patch to use the sys_*() functions directly.
drivers/gpu/drm/qxl This one uses a worker as a buffer between the mmap damage tracking and fb_*() functions, and flushing of the changes. I'll give it a go.
Studying these in detail and looking at the git log was useful as it showed me that the (struct fb_ops *)->fb_*() functions can be called in interrupt context. This means that I need the irq version of spin_lock(). It also validates my choice to defer these calls using the mmap defer damage worker ((struct fb_info).deferred_work), because it ensures that the dirty() call will always run in process context.
Noralf.